{"conference":{"name":"ISMIR 2026","date":"2026 Nov 8-12","timezone":"Asia/Dubai","miniconfUrl":"https://ismir2026program.ismir.net","officialSiteUrl":"https://ismir2026.ismir.net/","helpdeskUrl":"https://ismir2026.slack.com/archives/C0C45QQMU6A","venueCity":"Abu Dhabi","slackAppName":"ISMIR2026 Lounge"},"days":[{"day":"Mon","date":"2026-11-09","sessions":["PS-01","PS-02","PS-03","PS-04"]},{"day":"Tue","date":"2026-11-10","sessions":["PS-05","PS-06","PS-07","PS-08"]},{"day":"Wed","date":"2026-11-11","sessions":["PS-09","PS-10","PS-11","PS-12"]},{"day":"Thu","date":"2026-11-12","sessions":["PS-13","PS-14"]}],"sessions":{"PS-01":{"id":"PS-01","day":"Mon","date":"2026-11-09","start":"10:00","end":"11:30","slots":[{"kind":"oral","date":"2026-11-09","start":"10:00","end":"10:30"},{"kind":"poster","date":"2026-11-09","start":"10:30","end":"11:30"}],"papers":["363","7","12","26","524","28","538","94","105","114"]},"PS-02":{"id":"PS-02","day":"Mon","date":"2026-11-09","start":"11:30","end":"13:00","slots":[{"kind":"oral","date":"2026-11-09","start":"11:30","end":"12:00"},{"kind":"poster","date":"2026-11-09","start":"12:00","end":"13:00"}],"papers":["126","129","139","148","223","225","229","230","240","247"]},"PS-03":{"id":"PS-03","day":"Mon","date":"2026-11-09","start":"14:00","end":"15:30","slots":[{"kind":"oral","date":"2026-11-09","start":"14:00","end":"14:30"},{"kind":"poster","date":"2026-11-09","start":"14:30","end":"15:30"}],"papers":["98","33","34","49","51","56","72","3","112","120"]},"PS-04":{"id":"PS-04","day":"Mon","date":"2026-11-09","start":"16:30","end":"18:00","slots":[{"kind":"oral","date":"2026-11-09","start":"16:30","end":"17:00"},{"kind":"poster","date":"2026-11-09","start":"17:00","end":"18:00"}],"papers":["35","78","80","146","218","370","403","409","418","444"]},"PS-05":{"id":"PS-05","day":"Tue","date":"2026-11-10","start":"10:00","end":"11:30","slots":[{"kind":"oral","date":"2026-11-10","start":"10:00","end":"10:30"},{"kind":"poster","date":"2026-11-10","start":"10:30","end":"11:30"}],"papers":["371","261","267","271","251","318","321","424","459","485"]},"PS-06":{"id":"PS-06","day":"Tue","date":"2026-11-10","start":"11:30","end":"13:00","slots":[{"kind":"oral","date":"2026-11-10","start":"11:30","end":"12:00"},{"kind":"poster","date":"2026-11-10","start":"12:00","end":"13:00"}],"papers":["164","184","187","188","192","279","196","198","199","201"]},"PS-07":{"id":"PS-07","day":"Tue","date":"2026-11-10","start":"14:00","end":"15:30","slots":[{"kind":"oral","date":"2026-11-10","start":"14:00","end":"14:30"},{"kind":"poster","date":"2026-11-10","start":"14:30","end":"15:30"}],"papers":["202","237","250","256","266","195","280","534","293","523"]},"PS-08":{"id":"PS-08","day":"Tue","date":"2026-11-10","start":"15:30","end":"17:00","slots":[{"kind":"oral","date":"2026-11-10","start":"15:30","end":"16:00"},{"kind":"poster","date":"2026-11-10","start":"16:00","end":"17:00"}],"papers":["448","458","468","315","317","384","342","351","303","372"]},"PS-09":{"id":"PS-09","day":"Wed","date":"2026-11-11","start":"09:00","end":"10:30","slots":[{"kind":"oral","date":"2026-11-11","start":"09:00","end":"09:30"},{"kind":"poster","date":"2026-11-11","start":"09:30","end":"10:30"}],"papers":["205","13","477","128","141","182","357","4","428","74"]},"PS-10":{"id":"PS-10","day":"Wed","date":"2026-11-11","start":"11:30","end":"13:00","slots":[{"kind":"oral","date":"2026-11-11","start":"11:30","end":"12:00"},{"kind":"poster","date":"2026-11-11","start":"12:00","end":"13:00"}],"papers":["490","494","509","511","75","208","246","287","84","338"]},"PS-11":{"id":"PS-11","day":"Wed","date":"2026-11-11","start":"14:00","end":"15:30","slots":[{"kind":"oral","date":"2026-11-11","start":"14:00","end":"14:30"},{"kind":"poster","date":"2026-11-11","start":"14:30","end":"15:30"}],"papers":["373","379","386","397","405","406","87","289","206","309"]},"PS-12":{"id":"PS-12","day":"Wed","date":"2026-11-11","start":"15:30","end":"17:00","slots":[{"kind":"oral","date":"2026-11-11","start":"15:30","end":"16:00"},{"kind":"poster","date":"2026-11-11","start":"16:00","end":"17:00"}],"papers":["413","421","422","493","520","270","322","327","345","402"]},"PS-13":{"id":"PS-13","day":"Thu","date":"2026-11-12","start":"10:00","end":"11:30","slots":[{"kind":"oral","date":"2026-11-12","start":"10:00","end":"10:30"},{"kind":"poster","date":"2026-11-12","start":"10:30","end":"11:30"}],"papers":["512","39","60","83","452","258","269","298","433","414"]},"PS-14":{"id":"PS-14","day":"Thu","date":"2026-11-12","start":"11:30","end":"13:00","slots":[{"kind":"oral","date":"2026-11-12","start":"11:30","end":"12:00"},{"kind":"poster","date":"2026-11-12","start":"12:00","end":"13:00"}],"papers":["125","525","527","539","552","335","472","518","530","320"]}},"papers":{"363":{"uid":"363","title":"Learning Music Style for Piano Arrangement Through Cross-Modal Bootstrapping","authors":"Jingwei Zhao; Gus Xia; Ziyu Wang; Ye Wang","abstract":"What is music style? Though often described using text labels such as \"swing,\" \"classical,\" or \"emotional,\" the real style remains implicit and hidden in concrete music examples. In this paper, we introduce a cross-modal framework that learns implicit music styles from raw audio and applies them to symbolic music generation. Inspired by BLIP-2, our model leverages a Querying Transformer (Q-Former) to extract style representations from a large, pre-trained audio language model (LM), and further applies them to condition a symbolic LM for generating piano arrangements. We adopt a two-stage training strategy: contrastive learning to align auditory style with symbolic expression, followed by generative modeling for music arrangement. Our model generates piano performances jointly conditioned on a lead sheet (content) and a reference audio example (style), enabling controllable and stylistically faithful arrangement. Experiments demonstrate the effectiveness of our approach in piano cover generation, style transfer, and audio-to-MIDI retrieval, achieving substantial improvements in style-aware alignment and music quality.","primarySubject":"Generative Tasks","secondarySubject":"Applications -> music composition, performance, and production;Creativity -> creative practice involving MIR or generative technology;Generative Tasks -> music and audio synthesis;Musical features and properties -> musical style and genre;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p1-1-learning-music-style","channelUrl":"https://ismir2026.slack.com/archives/C0C5H63UBQX","funFact":"Jingwei used to honour British spelling, but in writing this paper he honors American spelling.\nDr. Ziyu Wang will be joining MBZUAI as a professor in 2027. Congrats, Ziyu, and all the best!\nThis paper was first submitted to ISMIR 2025 and got rejected. However, thanks to the reviewers' comments, we were able to substantially improve the experiments and get it accepted this year. (Yay!)","flags":[],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_363.html","displayId":"P1-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Applications","Music composition, performance, and production","Creativity","Creative practice involving MIR or generative technology","Music and audio synthesis","Musical features and properties","Musical style and genre","Representation learning","Open Review"]},"7":{"uid":"7","title":"Dual-Time Representation Framework for Interactive Music Transcription","authors":"Yu Sugimoto; Ryo Nishikimi; Eita Nakamura","abstract":"We present a framework for interactive music transcription (IMT) that enables users to efficiently correct outputs of automatic music transcription (AMT) methods. While existing IMT systems typically use physical time (PT) representations (measured in seconds), which can precisely describe note timings, metrical time (MT) representations (measured in fractions of a beat) are more suitable for score typesetting and can reduce the user’s cognitive cost in understanding musical structure. To enable corrections in MT-based note sequences (MTNSs) derived from PT-based note sequences (PTNSs) produced by AMT methods, and to support coherent and flexible operations across both time domains, we propose data structures consisting of beat annotations, a MTNS, and a PTNS. Moreover, to facilitate user interaction, we develop three core processing methods: rhythm quantization (PTNS to MTNS), onset/offset time alignment (MTNS to PTNS), and candidate generation for note corrections. These methods are formulated within a unified generative model that incorporates statistical knowledge of musical scores and performances. Taking singing transcription as a concrete example, we implemented an IMT system based on the proposed framework, and experimentally examined its potential in improving the human correction process.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Cognitive and user-centered MIR -> human-computer interaction","presentation":"In-person","slackChannel":"p1-2-dual-time-representation","channelUrl":"https://ismir2026.slack.com/archives/C0C6STF75NU","funFact":"","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_7.html","displayId":"P1-02","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Cognitive and user-centered MIR","Human-computer interaction","Open Review"]},"12":{"uid":"12","title":"TUTTI: Toward generalizable audio-to-score transcription via fully synthesized data","authors":"Jianhuai Hu; Yashan Wang; Shangda Wu; Zhancheng Guo; Shijie Liang; Wuna Meng; Chuanqi Yang; Xiaobing Li; Feng Yu; Maosong Sun","abstract":"Generalizable Audio-to-Score (A2S) transcription is fundamentally constrained by the severe scarcity of high-quality, real-world paired data. Relying solely on existing human-annotated datasets often restricts the generalization of A2S models, limiting their efficacy primarily to single-instrumentation domains. To break this dependency on scarce real-world data, we introduce TUTTI (Transformer for Unified audio-To-score Transcription trained on Synthetic multi-Instrumentation Data), a pre-training paradigm driven by a purely synthetic, large-scale dataset. Rather than using human-composed scores, we leverage a symbolic music generation model to generate a massive, highly scalable multi-instrumentation corpus and create audio-score pairs with expressive acoustic characteristics. Capitalizing on the generated data, we employ a standard Transformer encoder-decoder architecture. We empirically demonstrate that pre-training a unified attention-based model on generated, multi-instrumentation data yields a consistently stronger foundational representation than single-instrumentation training. When fine-tuned with downstream real-world datasets, TUTTI outperforms previous approaches, establishing new overall state-of-the-art results across various A2S baselines. Notably, TUTTI shows remarkable cross-instrument transferability, effectively adapting to unseen instruments with highly competitive performance. The source code and the TuttiCorpus dataset will be made publicly available at https://github.com/a-musiclover/TUTTI.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"In-person","slackChannel":"p1-3-tutti-toward-generalizable","channelUrl":"https://ismir2026.slack.com/archives/C0C6276QRFE","funFact":"I graduated as a chemist from the Univ. of Utah, a statistician from the U of South Carolina. Currently struggling at Central Conservatory of Music for my AI in music degree. But,  my love is piano.","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_12.html","displayId":"P1-03","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Open Review"]},"26":{"uid":"26","title":"Learning to Sing with Your Own Voice: Designing and Evaluating Generative AI-Assisted Vocal Coaching","authors":"Tianyu Gao; Shuqi Dai; Roger B. Dannenberg; Zhiyong Huang; Jin Song Dong; Jocelyn Dueck","abstract":"Traditional vocal coaching centers on vocal demonstration, yet voice mismatch between coach and student has long constrained its pedagogical effectiveness. Recent advances in generative AI (GenAI) make it feasible to generate personalized demonstrations in the student's own voice, potentially narrowing this gap. However, it remains unexplored how such demonstrations can be integrated into real-world coaching practice to augment human coaching. In this paper, we present the first controlled real-world study of GenAI-assisted vocal coaching, a new form of coaching in which AI demos in the student's own voice serve as the only source of vocal demonstration while the human coach provides verbal and interpretive guidance. We evaluate the feasibility of this approach with 22 student singers, comparing single-session GenAI-assisted coaching with traditional human coaching under blind expert assessment. The results show that the GenAI-assisted group achieved immediate technical and perceptual gains comparable to expert human coaching. Student and coach feedback further revealed a shift in coaching dynamics, with students gaining greater autonomy, despite an imperfect connection to the cloned voice. These findings provide the first empirical evidence that GenAI-assisted coaching can augment, rather than replace, human coaching in vocal pedagogy, and we derive design implications for future MIR systems supporting GenAI-assisted vocal coaching.","primarySubject":"Applications -> music training and education","secondarySubject":"Cognitive and user-centered MIR -> personalization;Cognitive and user-centered MIR -> user-centered evaluation;Creativity -> creative practice involving MIR or generative technology;Creativity -> creativity and learning;Creativity -> human-AI co-creativity","presentation":"In-person","slackChannel":"p1-4-learning-to-sing","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5RMCBH","funFact":"Vocal coaching is fine!","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_26.html","displayId":"P1-04","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Applications","Music training and education","Cognitive and user-centered MIR","Personalization","User-centered evaluation","Creativity","Creative practice involving MIR or generative technology","Creativity and learning","Human-AI co-creativity","Open Review"]},"524":{"uid":"524","title":"Vocal Music under Phoneme-Conditional Analysis","authors":"Hayoon Kim; Kyogu Lee","abstract":"The vocal music of each language carries a distinctive sonic identity, even without instrumental accompaniment. We ask whether these differences are measurable and traceable to specific phonemes. To tackle this question, we introduce phoneme-conditional analysis, which isolates the acoustic effect of typologically distinctive phonemes by comparing marker syllables against matched non-marker controls within the same song, holding singer, melody, and genre constant. Across nine typologically diverse languages and thousands of songs, we measure effects along five acoustic dimensions. Song-level profiles built from these effects identify the language of an unaccompanied vocal at 85.5% balanced accuracy in a nine-way classification with folds grouped by artist; whether the separability arises by accumulation of the phoneme-local effects themselves is left open. Our findings suggest that phonological structure leaves systematic and measurable traces in how each language is sung.","primarySubject":"Computational musicology -> computational ethnomusicology","secondarySubject":"MIR data fundamentals -> lyrics and other textual data","presentation":"In-person","slackChannel":"p1-5-vocal-music-under","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6QEBJT","funFact":"Hayoon Kim learned Arabic to attend ISMIR.\nHayoon Kim was too broke to fly to Abu Dhabi.\nHayoon Kim dreams of eating dates beneath a palm tree beside a camel.","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_524.html","displayId":"P1-05","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Computational musicology","Computational ethnomusicology","MIR data fundamentals","Lyrics and other textual data"]},"28":{"uid":"28","title":"Pitch Estimation and Vocal Melody Extraction with Self-supervised Phase Comparators","authors":"Li Su","abstract":"Pitch estimation based on self-repetition measures is a classical yet theoretically well-grounded approach, offering the potential to address specific challenges overlooked by current supervised and self-supervised pitch detection methodologies. In this work, we employ Self-Supervised Learning (SSL) to directly learn the time-domain similarity of signals for joint estimation of pitch and pitch confidence scores. We frame this task as a problem of training phase comparators, and introduce magnitude-phase swapping as a novel data augmentation technique for training. Experiments on pitch estimation and vocal melody extraction demonstrate that the proposed representation achieves performance comparable to that of state-of-the-art SSL-based models. Finally, we explore the special characteristics of such representations in isolating individual note segments.","primarySubject":"MIR data fundamentals -> music signal processing","secondarySubject":"Analysis tasks -> music transcription and annotation","presentation":"","slackChannel":"p1-6-pitch-estimation-and","channelUrl":"https://ismir2026.slack.com/archives/C0C6STFGNL8","funFact":"This research was inspired by the ancient philosophical idea that “You cannot step twice into the same river, for it is not the same river and you are not the same you.”","flags":[],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_28.html","displayId":"P1-06","presenterMode":"","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Music signal processing","Analysis tasks","Music transcription and annotation","Open Review"]},"538":{"uid":"538","title":"Real-Time Interactive Music Generation via Training-Data-Free Streaming Consistency Distillation","authors":"Baisen Wang; Chenxi Bao; Qisong Han","abstract":"Modern music generators remain largely confined to high-latency offline rendering. We introduce training-data-free streaming consistency distillation, which converts a pre-trained text-to-music model into a chunk-wise autoregressive instrument. Prompt-only teacher rollouts eliminate paired target audio and precomputed latent trajectories. A one-step student follows the teacher’s cached trajectory, while latent, spectral, and temporal-difference objectives preserve musical detail under aggressive acceleration. The system continuously extends a latent stream and accepts semantic control updates without restarting playback. On an NVIDIA H200, it reaches 0.086 s first-chunk latent latency and a 0.009 real-time factor. Objective and subjective evaluations show improved fidelity over latent-only streaming and responsive phrase-level steering, reframing text-to-music generation as a playable medium for human–AI co-creation.","primarySubject":"Creativity -> human-AI co-creativity","secondarySubject":"Generative Tasks -> music and audio synthesis;Generative Tasks -> real-time considerations","presentation":"In-person","slackChannel":"p1-7-real-time-interactive","channelUrl":"https://ismir2026.slack.com/archives/C0C5WK084A2","funFact":"When he isn't distilling streaming music models, Baisen Wang works as an IT teacher at The High School Affiliated to Renmin University of China—where his students put \"real-time human-AI interaction\" to the test.","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_538.html","displayId":"P1-07","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Creativity","Human-AI co-creativity","Generative Tasks","Music and audio synthesis","Real-time considerations"]},"94":{"uid":"94","title":"Toward Song Exploration Without Listening Through Lyrics-Derived Linguistic Information","authors":"Kanako Suzuki; Kosetsu Tsukuda; Takayuki Nakatsuka; Kento Watanabe; Tomoyasu Nakano; Masataka Goto; Keitaro Tanaka; Shigeo Morishima","abstract":"Music streaming services typically assume that users explore songs while listening to them, which often exposes users to songs that do not match their preferences. Although lyrics are an important cue for judging songs, reading the full lyrics of many songs during exploration imposes substantial cognitive load. In this paper, we investigate whether lyrics-derived linguistic information can support song exploration without listening. Specifically, we use a large language model to generate 17 types of linguistic information from song lyrics, including a catchphrase, a motif, and a summary text, and conduct a crowdsourcing-based user study involving 1,235 participants. Our analysis results provide several reusable insights, including the following: (1) There is substantial demand for exploring songs by examining various kinds of information so as to avoid listening to songs that do not match users' preferences. (2) When lyrics-derived linguistic information is judged as \"Dislike,\" the full lyrics are also often judged as \"Dislike.\" (3) Combining multiple types of linguistic information increases preference consistency with the full lyrics while requiring much less text than reading the full lyrics. We also discuss how these insights can be applied to song exploration systems.","primarySubject":"Cognitive and user-centered MIR -> user-centered evaluation","secondarySubject":"Applications -> music retrieval systems;MIR data fundamentals -> lyrics and other textual data","presentation":"Virtual","slackChannel":"p1-8-toward-song-exploration","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA6C42F","funFact":"The idea for this research came from the first author’s own struggle to find new favorite songs without having to listen to songs that might not match her preferences.\nKanako Suzuki practiced ballet for 13 years and later spent two years learning locking dance.\nMusic has always been a big part of her life, and she never goes a day without singing.","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_94.html","displayId":"P1-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","User-centered evaluation","Applications","Music retrieval systems","MIR data fundamentals","Lyrics and other textual data"]},"105":{"uid":"105","title":"FMANet: Efficient and Lightweight Beat Tracking via Fourier Modulation Attention","authors":"Ganghui Ru; Yi Yu; Wei Li","abstract":"Achieving high-precision beat and downbeat tracking while maintaining model compactness remains a significant challenge in music information retrieval. \nAlthough recent state-of-the-art models deliver impressive results, they often rely on massive architectures, leading to substantial computational redundancy. \nIn this paper, we propose FMANet, an efficient end-to-end framework for joint beat and downbeat tracking. \nOur design is centered on the Fourier Modulation Attention (FMA), which is motivated by the inherent periodicity of musical rhythm. \nBy invoking the spectrum convolution theorem, FMA enables the model to capture long-range rhythmic dependencies through differentiable frequency-domain modulations, achieving a global receptive field with linear-logarithmic complexity. \nTo complement this global modeling, we introduce a Multi-scale Feed-Forward Network (MSFFN) to refine local harmonic features and ensure precise beat localization. Experimental results on seven public datasets demonstrate that FMANet achieves performance comparable to state-of-the-art architectures while utilizing only 1.6M parameters. \nThis significant reduction in complexity, achieved without sacrificing tracking accuracy, underscores FMANet’s potential as a specialized and lightweight solution for rhythmic structure analysis.","primarySubject":"Musical features and properties -> rhythm, beat, tempo","secondarySubject":"MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p1-9-fmanet-efficient-and","channelUrl":"https://ismir2026.slack.com/archives/C0C60EZBFNV","funFact":"","flags":["student"],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_105.html","displayId":"P1-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Rhythm, beat, tempo","MIR data fundamentals","Music signal processing"]},"114":{"uid":"114","title":"On the Geometry of Music Bandwidth Extension in Latent Spaces of Audio Codecs","authors":"Hendrik Vincent Koops; Hao Hao Tan; Elio Quinton","abstract":"Recent audio restoration increasingly relies on large-scale conditional latent generative modeling, including diffusion, Schrodinger Bridges, and Flow Matching variants, to invert degradations such as bandwidth limitation or noise.\nWe present an analysis of the performance of various state-of-the-art methods compared to simple arithmetic transformations in the latent spaces of multiple neural codecs for musical bandwidth extension. \nWe show that estimating a single transport vector between the clean and degraded latent centroids on a reference set, and adding it to degraded latents, can yield restoration performance competitive with large diffusion models. \nThis suggests, first, that some neural codec latent spaces exhibit structure aligned with audio bandwidth; and second, that in such cases complex conditional models may offer only limited gains over a simple vector addition.\nWe argue that these findings reveal an interesting avenue for future research whereby models could take advantage of the latent space structure in order to offer greater training and parameter efficiency, and overall better performance. \nAdditionally, we propose to consider this simple arithmetic transformation as a baseline for music bandwidth extension research, as it allows an assessment of the contribution of learnable parameters towards restoration performance.","primarySubject":"Generative Tasks -> transformations","secondarySubject":"MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p1-10-on-the-geometry","channelUrl":"https://ismir2026.slack.com/archives/C0C60EXHRG9","funFact":"","flags":[],"session":"PS-01","posterUrl":"https://ismir2026program.ismir.net/poster_114.html","displayId":"P1-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Transformations","MIR data fundamentals","Music signal processing","Open Review"]},"126":{"uid":"126","title":"Rethinking Automatic Music Mixing as Sequential Stem Blending","authors":"Yen-Tung Yeh; Chung-Jui Chan; Yun-Ning Hung; Yi-Hsuan Yang","abstract":"Automatic music mixing, the task of automatically combining individual audio tracks into a cohesive mixture, is typically addressed by parallelized architectures that process all input tracks in a single pass. In this work, inspired by how human mix engineers process stems one at a time, we propose a paradigm shift and ask whether automatic music mixing can be reformulated as a sequential stem blending task, where each stem is blended into a growing submix. Specifically, we train a latent flow matching model conditioned on the submix context, enabling sequential processing of an arbitrary number of input tracks. To train the model, we introduce a degradation-based data synthesis strategy that simulates realistic stem blending scenarios from existing multitrack and source separation datasets. Experimental results on both stem blending and automatic music mixing benchmarks demonstrate the effectiveness of the proposed approach. We provide audio examples on the accompanying demo page.","primarySubject":"Applications -> music composition, performance, and production","secondarySubject":"Generative Tasks -> artistically-inspired generative tasks;MIR data fundamentals -> music signal processing;Musical features and properties -> musical style and genre","presentation":"Virtual","slackChannel":"p2-1-rethinking-automatic-music","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7UV413","funFact":"1. There are more and more audio effects/mixing papers in this year!","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_126.html","displayId":"P2-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications","Music composition, performance, and production","Generative Tasks","Artistically-inspired generative tasks","MIR data fundamentals","Music signal processing","Musical features and properties","Musical style and genre","Open Review"]},"129":{"uid":"129","title":"Song Aesthetics Evaluation with Multi-Stem Attention and Hierarchical Uncertainty Modeling","authors":"Yishan Lv; Jing Luo; Boyuan Ju; Yang Zhang; Xinda Wu; Bo Yuan; Xinyu Yang","abstract":"Music generative artificial intelligence (AI) is rapidly expanding music content, necessitating automated song aesthetics evaluation for efficient music retrieval and recommendation. However, existing studies largely focus on speech, audio or singing quality, leaving song aesthetics underexplored. Moreover, conventional approaches often predict a precise Mean Opinion Score (MOS) value directly, which struggles to capture the nuances of human perception in song aesthetics evaluation. This paper proposes a song-oriented aesthetics evaluation framework, featuring two novel modules: 1) Multi-Stem Attention Fusion (MSAF) builds bidirectional cross-attention between mixture-vocal and mixture-accompaniment pairs, fusing them to capture complex musical features; 2) Hierarchical Granularity-Aware Interval Aggregation (HiGIA) learns multiple granularity score probability distributions, aggregates them into a score interval, and applies a regression within the interval to produce the final score. We evaluated on two datasets of full-length songs: SongEval dataset (AI-generated) and an internal aesthetics dataset (human-created), and compared with two state-of-the-art (SOTA) models. Results show that the proposed method achieves stronger performance for multi-dimensional song aesthetics evaluation.","primarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology","secondarySubject":"Cognitive and user-centered MIR -> user-centered evaluation;MIR data fundamentals -> music signal processing;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"Virtual","slackChannel":"p2-2-song-aesthetics-evaluation","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCA5802","funFact":"","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_129.html","displayId":"P2-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Evaluation methodology","Cognitive and user-centered MIR","User-centered evaluation","MIR data fundamentals","Music signal processing","Musical features and properties","Timbre, instrumentation, and singing voice"]},"139":{"uid":"139","title":"Exploring the Design Space of Representation Learning for Audio Transformations","authors":"Sungho Lee; Marco A. Martínez-Ramírez; Junghyun Koo; Wei-Hsiang Liao; Kyogu Lee; Yuki Mitsufuji","abstract":"Neural audio representation learning has enabled a range of content-oriented applications, but the resulting features remain limited for tasks involving audio processing. Furthermore, it is not obvious what processing-aware representations should capture: the processing itself, abstracted away from source content, or the processed audio that retains it. Existing approaches implicitly commit to one or the other and also differ in their models, data, and evaluation, obscuring which design choices drive their behavior. We address both questions within a unified framework of three objectives: processing consistency, description alignment, and equivariance via forward prediction. We compare all combinations of the objectives under a controlled setup and reveal their relative strengths and interactions. Our framework produces both a transformation embedding and a processed-audio embedding, and we find that the two play complementary roles: distance-based tasks favor the former, while probe-based tasks favor the latter. Combined with improvements in network architecture and training pipeline, our representations outperform prior baselines across retrieval, probe-based evaluation, and style transfer.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks -> similarity metrics;MIR data fundamentals -> music signal processing;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p2-3-exploring-the-design","channelUrl":"https://ismir2026.slack.com/archives/C0C627A8Q0L","funFact":"","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_139.html","displayId":"P2-03","presenterMode":"online","materials":["paper"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Similarity metrics","MIR data fundamentals","Music signal processing","Representations of music","Open Review"]},"148":{"uid":"148","title":"UOT-IR: Structured Routing of High-Polyphony Symbolic Music into Fixed-Budget Representations","authors":"Ziyue Kang; Nan Nan; Chenhao Lin; Xiaohong Guan","abstract":"High-polyphony symbolic music is increasingly used in generation, analysis, and arrangement, yet many downstream tasks require bounded representations with fixed tracks or slots. Converting richly orchestrated scores into compact forms is therefore necessary, but existing approaches relying on heuristic simplification or generic representation-space reduction often fail to preserve structural roles, orchestration compatibility, and playability under strict budgets. To address the issue, this study reformulates the compression problem as a fixed-budget structured routing problem and proposes Unbalanced Optimal Transport for Information Routing (UOT-IR), a training-free framework based on constrained unbalanced optimal transport. UOT-IR combines an orchestration prior, adaptive marginal relaxation, temporal decoding, and playability-aware projection to produce compact and musically coherent bounded representations. This work further studies two practical settings under the same slot budget: template standardization, which maps each input to a predefined bounded template, and adaptive preservation, which retains representative content without assuming an external template. Experiments on the SymphonyNet corpus show that UOT-IR delivers strong overall performance across both settings, including the best Note-F1 in adaptive preservation (0.9120), together with the lowest structural cost (14.7165) and bad structural confusion rate (0.3406) in template standardization. This work establishes a principled paradigm for fixed-budget symbolic music compression, offering a practical path toward compact, structured, and musically coherent symbolic representations.","primarySubject":"MIR data fundamentals -> symbolic music processing","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation metrics;Musical features and properties -> representation learning;Musical features and properties -> representations of music;Musical features and properties -> structure, segmentation, and form;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"Virtual","slackChannel":"p2-4-uot-ir-structured","channelUrl":"https://ismir2026.slack.com/archives/C0C5YC9LWE6","funFact":"","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_148.html","displayId":"P2-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Symbolic music processing","Evaluation, datasets, and reproducibility","Evaluation metrics","Musical features and properties","Representation learning","Representations of music","Structure, segmentation, and form","Timbre, instrumentation, and singing voice","Open Review"]},"223":{"uid":"223","title":"From Prompting to Describing: A Cross-Cultural Study of Language for AI-Generated Music","authors":"Sangheon Park; Claire Arthur","abstract":"Text-to-music (TTM) generation systems allow users to create music through natural language prompts, yet it is unclear whether the descriptive language used to prompt aligns with descriptive language used to summarize or describe heard music. We pair 200 real-world Udio prompts with their generated audio and free-form descriptions collected from English- n = 70 and Korean-speaking n = 78 listeners, and contribute a human-derived taxonomy of musical prompting vocabulary grounded in real user data. Using this framework, alongside word- and vector-level analyses, we find a consistent structural asymmetry: prompts are dominated by Genre and Story/Narrative language. Genre terms propagate most reliably from prompt to perception, while narrative-heavy prompts are the strongest predictor of semantic misalignment. A preliminary cross-cultural comparison further suggests that description profiles vary across listener populations along narrative, functional, and affective dimensions, raising questions about whether current TTM systems, trained on aggregated English-centric corpora, can accommodate the full diversity of how people naturally express musical ideas.","primarySubject":"Cognitive and user-centered MIR -> user-centered evaluation","secondarySubject":"Cognitive and user-centered MIR -> human-computer interaction;Cognitive and user-centered MIR -> perception and cognition;MIR data fundamentals -> lyrics and other textual data;MIR data fundamentals -> web mining, and natural language processing","presentation":"Virtual","slackChannel":"p2-5-from-prompting-to","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5X8G15","funFact":"For IRB approval, Sangheon had to manually listen to and review every Udio track to make sure there were no harassing or sexually explicit words. I even missed one track, which caused the IRB submission to be rejected once.","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_223.html","displayId":"P2-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","User-centered evaluation","Human-computer interaction","Perception and cognition","MIR data fundamentals","Lyrics and other textual data","Web mining, and natural language processing","Open Review"]},"225":{"uid":"225","title":"Phase-Based Onset Enhancement for Piano Transcription in Ensemble Settings","authors":"Hyemi Kim; Jiyun Park; Juhan Nam","abstract":"Piano transcription models have achieved very high performance in recent state-of-the-art systems, but their performance tends to degrade in ensemble settings where the piano is played alongside other instruments.\nWhile some approaches attempt to address this issue by incorporating multi-instrumental datasets, the problem remains challenging. \nIn particular, when other instruments produce high-energy sustained tones while relatively low-energy piano notes occur simultaneously, the magnitude components of their frequency representations overlap, making it difficult to accurately distinguish piano notes.\nTo address this issue, we propose an improved piano transcription model that jointly leverages magnitude and phase information, enabling more robust piano transcription performance in ensemble settings. \nWe evaluate the proposed model on two piano ensemble datasets with precise MIDI note annotations, which we obtained by revisiting and extending existing datasets.\nNotably, we constructed a new piano concerto dataset, PCD-K, by playing a piano part on a Disklavier and recording the audio and MIDI pairs with pre-recorded orchestral tracks. \nExperimental results demonstrate that phase information is highly effective in improving performance, outperforming magnitude-based approaches in ensemble settings.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"Virtual","slackChannel":"p2-6-phase-based-onset","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5WB29M","funFact":"The researchers in Juhan Nam’s lab, along with Prof. Nam himself, became the performers, too, playing the piano parts of piano concertos to build the PCD-K dataset.\nAs mostly amateur pianists, practicing these parts was almost as challenging as the research itself.\nSpecial thanks to Yonghyun Kim, then at KAIST and now at Georgia Tech, for taking on many of the challenging pieces.","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_225.html","displayId":"P2-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Open Review"]},"229":{"uid":"229","title":"Online Music Concerts Leveraging Existing Music Videos: A Case Study on a Music Web Service","authors":"Kosetsu Tsukuda; Keisuke Ishida; Masahiro Hamasaki; Masataka Goto","abstract":"The videos streamed during online music concerts (OMCs) are generally videos created specifically for OMCs or live broadcasts of artists' performances. This entails significant time and cost to organize OMCs. In this paper, we present an analytical case study on OMCs that can be hosted at lower cost by leveraging existing music videos. In the OMC format, participants can react to the music as they would at an in-person music concert (IMC), and these reactions are visible to other participants. We analyzed user behavior logs from 72 OMCs hosted by various organizers on a web service, involving a total of 20,579 participants. Our analysis yielded valuable insights, including the following: (1) The majority of participants, once they joined an OMC, tended to stay for long periods, demonstrating the perceived value of participating in OMCs using existing music videos. (2) The OMC format also provided an opportunity for participants to discover new music of interest. (3) Participants actively engaged in behaviors similar to those observed during IMCs by using the provided features to express their reactions to music.","primarySubject":"Applications","secondarySubject":"","presentation":"Virtual","slackChannel":"p2-7-online-music-concerts","channelUrl":"https://ismir2026.slack.com/archives/C0C6STL436U","funFact":"Kosetsu Tsukuda specializes in recommender systems and user behavior analysis based on large-scale surveys and log data.\nKosetsu Tsukuda has presented seven first-authored papers at ISMIR: one each in 2017, 2018, and 2020, and two each in 2021 and 2023.\nYou might recognize Kiite Cafe from ISMIR 2021, where Kosetsu Tsukuda and colleagues first presented the service in their paper, \"Kiite Cafe: A Web Service for Getting Together Virtually to Listen to Music.\"","flags":[],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_229.html","displayId":"P2-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications"]},"230":{"uid":"230","title":"VioLM: A Neural Language Model for Violin Synthesis with Articulation","authors":"Sunan Zou; Zhe Zhang; Yigitcan Özer; Guojie Luo; Junichi Yamagishi","abstract":"Music performance synthesis (MPS) for musical instruments remains challenging due to insufficient expressive datasets and limited modeling of instrument-specific playing techniques. Instruments such as the violin require handling multiple articulations with distinct physical excitation mechanisms, yet existing neural network-based MPS models largely ignore such distinctions. We propose VioLM, a language model (LM)-based violin synthesis framework that generates expressive audio from MIDI scores with articulation labels. To address data scarcity, VioLM combines an expressive data-generation pipeline with a two-stage transfer-learning strategy. We propose the Capriccio synthetic violin dataset, built from large-scale MIDI corpora with velocity simulation, articulation annotation, and high-quality audio rendering. Two-stage transfer learning first adapts a pre-trained piano VALL-E model to the violin domain using Capriccio then enhances audio fidelity via LoRA fine-tuning on real recordings. Experiments show VioLM achieves strong subjective and objective performance as the first neural MPS system with explicit violin articulation control.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"In-person","slackChannel":"p2-8-violm-a-neural","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJ60H38","funFact":"","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_230.html","displayId":"P2-08","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Evaluation, datasets, and reproducibility","Novel datasets and use cases"]},"240":{"uid":"240","title":"Genre Bias or Aesthetic Perception? Identifying and Mitigating Shortcut Learning in Music Evaluation","authors":"Yizhou Zhang; Wangjin Zhou; Yi Zhao; Wei Tan; Keisuke Imoto; Zhi Gong","abstract":"Music aesthetics scoring plays a critical role in applications such as dataset curation, generative model evaluation, and reward modeling for music generation. Recent approaches rely on deep neural networks trained on human-annotated ratings, but these models may exploit spurious correlations rather than capturing perceptually meaningful aesthetics. In this work, we identify a previously underexplored failure mode in music evaluation models: genre-induced shortcut learning. Through a systematic analysis of SongEval, we show that biases in training data lead to strong correlations between genre-related features and predicted scores, causing the model to use them as a proxy for aesthetics. This results in systematic overestimation of pop music and undervaluation of high-quality samples from other genres, leading to predictions that are inconsistent with human preferences. To address this issue, we propose a training objective that jointly reweights hard samples and regularizes group-level performance, encouraging the model to learn genre-invariant representations of musicality. Experimental results demonstrate that our method reduces genre-dependent bias and improves alignment with human preferences, as reflected by gains in both cross-genre and within-genre preference alignment.","primarySubject":"Generative Tasks -> evaluation metrics","secondarySubject":"Analysis tasks -> alignment, synchronization, and score following;Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> evaluation metrics;Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p2-9-genre-bias-or","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJ682Q6","funFact":"","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_240.html","displayId":"P2-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Evaluation metrics","Analysis tasks","Alignment, synchronization, and score following","Evaluation, datasets, and reproducibility","Evaluation methodology","Music and audio synthesis"]},"247":{"uid":"247","title":"Gugak-VocalSet: A Multi-Genre Studio-Quality Dataset of Korean Traditional Singing with Sigimsae Annotations","authors":"Danbinaerin Han; Wonil Kim; Seah Hong; Dongjoo Moon; Jaeok Lee; Jongpil Lee; Chaewon Kim; Juhan Nam","abstract":"We present Gugak-VocalSet, the first publicly available audio dataset for Korean traditional singing with dense annotations of sigimsae, a structured pitch contour system of vocal expressions. Gugak-VocalSet contains 24.5 hours of unaccompanied solo vocal recordings across 305 tracks of representative repertoire curated by the National Gugak Center, performed by 20 professional singers in a controlled studio environment with fixed key and tempo. In addition to genre, lyrics, and mood/timbre tags, the dataset provides a unified sigimsae ontology and approximately 68,000 segments (71,700 instances including multi-label) of audio-grounded sigimsae annotations; the ontology adopts a functional-position-based hierarchy designed for consistent application across genres, and 14 expert annotators specializing in Korean traditional music performed segment-level labeling accordingly. Baseline experiments on sigimsae classification and event detection show that domain-informed F0 representations remain competitive with, and on classification surpass, frozen probing of a general-purpose music foundation model, while pitch and spectral cues play complementary roles in event-level detection.","primarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","secondarySubject":"Computational musicology -> computational ethnomusicology;Musical features and properties -> expression and performative aspects of music","presentation":"Virtual","slackChannel":"p2-10-gugak-vocalset-a","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA94EUT","funFact":"1. Danbinaerin Han was in charge of the K-culture night at ISMIR 2025 in Daejeon. If you were there, please never forget that night.\n2.  Danbinaerin Han played the Haegeum (a Korean two-string fiddle) in the music program at her very first ISMIR in 2022, and has hardly had a chance to play it since.\n3. Curious about the data? Take a listen on the demo page, labels included: www.danbinaerin.com/Gugak-VocalSet/","flags":["student"],"session":"PS-02","posterUrl":"https://ismir2026program.ismir.net/poster_247.html","displayId":"P2-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Novel datasets and use cases","Computational musicology","Computational ethnomusicology","Musical features and properties","Expression and performative aspects of music","Open Review"]},"98":{"uid":"98","title":"Masked Diffusion Enables Coherent Beat Tracking","authors":"Francesco Foscarin; Filip Korzeniowski; Richard Vogl","abstract":"Current neural networks for beat tracking generate invalid outputs, such as consecutive downbeats and erratic tempo changes, even when these are not present in the training data.\nHeavy post-processing techniques can alleviate these problems, but the original cause of this inconsistent behaviour remains unknown.\nWe hypothesise that it stems from inadequate modelling of multiple plausible output beat grids, resulting in an invalid mixture of competing interpretations.\nWe propose a masked diffusion approach that properly models multiple outputs and enables the model to build coherent predictions through iterative inference.\nWe devise three modifications to standard masked diffusion that enable its application to beat tracking: independent masking of beats and downbeats during training and inference, a balanced masking scheduler for inference, and peak-picking across inference steps.\nOur approach reduces erratic behaviours and improves beat-tracking performance.","primarySubject":"Musical features and properties -> rhythm, beat, tempo","secondarySubject":"Analysis tasks -> music transcription and annotation;Analysis tasks -> pattern matching and detection","presentation":"Virtual","slackChannel":"p3-1-masked-diffusion-enables","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA6J9A7","funFact":"- The \"weird-activation\" problem solved in the paper has been taunting the author for many years\n- This system is based on the open-source Beat This paper, from the same first author","flags":[],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_98.html","displayId":"P3-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Rhythm, beat, tempo","Analysis tasks","Music transcription and annotation","Pattern matching and detection","Open Review"]},"33":{"uid":"33","title":"MuScriptor: An Open Model for Multi-Instrument Music Transcription","authors":"Simon Rouard; Michael Krause; Axel Roebel; Carl-Johann Simon-Gabriel; Alexandre Défossez","abstract":"Existing methods for automatic music transcription are often limited to single-instrument recordings or fail on complex, real music mixes. Although previous work utilizes synthetic training data, the resulting models generalize poorly, leading to largely unusable transcription output in realistic, multi-instrument settings.\nIn this work, we analyze the effectiveness of synthetic data for pre-training while combining it with fine-tuning on real music audio and post-training using reinforcement learning. We further introduce conditioning on instrument presence to customize transcriptions. Finally, we release MuScriptor, an open-weight multi-instrument music transcription model that works on real-world music recordings from across a diverse range of musical genres.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"","presentation":"","slackChannel":"p3-2-muscriptor-an-open","channelUrl":"https://ismir2026.slack.com/archives/C0C6277MR6G","funFact":"Michael Krause plays piano and should really practice more. Go tell him!\nSimon Rouard plays the drums in the band \"lost deposit\" that you hear in the demo at the end of the video!","flags":["student"],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_33.html","displayId":"P3-02","presenterMode":"","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Open Review"]},"34":{"uid":"34","title":"Pitch Shifting via Semantic Transformations in a Shared Audio-Symbolic Latent Space","authors":"Ben Maman; Andreas Brendel; Meinard Müller","abstract":"In this paper, we present a data-driven approach to pitch shifting based on a semantic transformation in a shared audio--symbolic latent space, rather than direct frequency scaling. We learn a multi-view representation that embeds audio together with complementary quasi-symbolic views such as pitch contours, piano rolls, and phonetic features, capturing both acoustic detail and structured musical information. Within this space, pitch-shift transformations are learned from symbolic training pairs obtained by transposing pitch-related quasi-symbolic views, avoiding the need for paired audio recordings at different pitch levels. The transformation is applied to embeddings derived from audio and decoded to generate pitch-shifted signals, with a subsequent fine-tuning stage improving reconstruction quality. Experiments on vocal--instrumental music demonstrate improved timbre preservation compared to a conventional signal-processing baseline, particularly for large pitch shifts, as well as substantially higher audio quality than a data-driven analysis--resynthesis approach. More broadly, the results suggest that semantic transformations in a shared audio--symbolic latent space provide a promising framework for controllable audio editing. Qualitative samples are available on our project page: https://latentpitchshift.github.io/latent-pitch-shift","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Generative Tasks -> music and audio synthesis;Generative Tasks -> transformations;Musical features and properties -> representations of music;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"Virtual","slackChannel":"p3-3-pitch-shifting-via","channelUrl":"https://ismir2026.slack.com/archives/C0C5YC6T9MY","funFact":"Ben is always glad to sing Umm Kulthum Songs","flags":[],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_34.html","displayId":"P3-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Generative Tasks","Music and audio synthesis","Transformations","Representations of music","Timbre, instrumentation, and singing voice","Open Review"]},"49":{"uid":"49","title":"Towards Robust Version Identification in the Wild: A Dataset, Benchmark, and Fine-Tuning Study","authors":"Simon Hachmeier; R. Oguz Araz; Dmitry Bogdanov; Robert Jäschke; Xavier Serra","abstract":"Existing datasets for musical version identification (VI) are primarily derived from curated metadata sources such as SecondHandSongs and Discogs, and are therefore dominated by professionally recorded tracks. This leads to a domain mismatch with real-world scenarios, where amateur and user-generated content is prevalent.\n\nTo address this limitation, we introduce DiVers, a large-scale VI dataset comprising over 1.1 million musical versions, with train–validation–test splits compatible with established datasets such as Discogs-VI-YT, SHS100K, and Da-TACOS. In addition to standard version-level annotations, DiVers provides automatically assigned tags (e.g., instrumental, live) and segment-level predictions indicating the presence or absence of music.\n\nWe evaluate the proposed dataset by training state-of-the-art VI systems. Our results show that models trained on DiVers achieve substantially improved robustness to acoustically diverse and noisy inputs, while maintaining a stable performance on cleaner, studio-quality benchmarks. We release the dataset metadata, code for its construction, and all experimental pipelines to support reproducibility.","primarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","secondarySubject":"Applications -> music retrieval systems;Musical features and properties -> representation learning","presentation":"In-person","slackChannel":"p3-4-towards-robust-version","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJ1S27Q","funFact":"- not trying to block YouTube videos with my research, but make them more easily findable :)\n- covered many songs on the electric guitar myself\n- my first ISMIR author paper, but attended already in San Francisco in 2024","flags":[],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_49.html","displayId":"P3-04","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Novel datasets and use cases","Applications","Music retrieval systems","Musical features and properties","Representation learning","Open Review"]},"51":{"uid":"51","title":"CrowdioSet and PaRIRset: Two Datasets Towards Live Music Source Separation","authors":"Enric Gusó; Xavier Serra","abstract":"Most Music Source Separation (MSS) models do not generalize well to live music recordings because they are trained on studio recordings alone, disregarding the venue acoustics, the speaker system's response and audience noise. We propose to bridge this gap by providing and training a model on two novel datasets. First, we present CrowdioSet: a noise dataset comprising 4800 real ambience tracks from Freesound and synthetic sing-alongs for the vocals in MUSDB18 and MOISESDB datasets, generated from zero-shot singing voice conversions. CrowdioSet enables effective audio denoising for live recordings, resulting in superior separation both in objective and subjective evaluations. Second, we introduce PaRIRset, a stereo impulse response dataset captured across 40 professional concert venues using a microphone array. Our results show that adding PaRIRset RIRs increases the performance of a MSS model compared to using real RIRs from Speech Enhancement tasks alone. We make the examples, code, model weights, PaRIRset, and CrowdioSet freely available to the public.","primarySubject":"Analysis tasks -> sound source separation","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"Virtual","slackChannel":"p3-5-crowdioset-and-parirset","channelUrl":"https://ismir2026.slack.com/archives/C0C6STFEE64","funFact":"PaRIRset RIRs were captured during the pre-show sound-checks of the last tours of Cala Vento, a well-known Spanish indie-rock band.","flags":["student"],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_51.html","displayId":"P3-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Sound source separation","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Open Review"]},"56":{"uid":"56","title":"Prima la Musica! Musically-Informed Symbolic Features for Section Boundary Detection in MIDI","authors":"ELVIO CIPOLLONE; FRANCESCO CIPOLLONE","abstract":"We present a section boundary detection system for symbolic music (MIDI) that achieves an F1 of 0.821 on the SLMS benchmark, improving over the previous state of the art by 4.2 percentage points under its evaluation protocol; a control experiment further quantifies how far transcription conventions in the corpus inflate absolute scores. Rather than learning boundary patterns from low-level representations, we encode specific musical observations about how section boundaries manifest. These observations (the texture reconfigures, the bar before a boundary thins out, the drummer announces the transition, melodic shapes break…) are translated into a set of 71 features organized in families, each corresponding to a distinct musical phenomenon. A 1D convolutional neural network ensemble whose kernel sizes align with the power-of-two phrase structure of Western music (2, 4, 8, and 16 bars) learns to combine these signals. The design of these features was itself the main source of improvement: the gains we measured during development consistently traced back to musical insights, like instrument identification heuristics and aligning kernel sizes with phrase structure. The most striking finding is that 26 explicit harmonic features, designed to encode chord identity and harmonic function, added no statistically significant contribution beyond raw pitch-class distributions and bass-line features, and were removed. We argue that in the symbolic domain, where pitch, rhythm, and instrumentation are directly observable, the most effective driver of improvement on this task was not model complexity but musical knowledge encoded as features","primarySubject":"Musical features and properties -> structure, segmentation, and form","secondarySubject":"Analysis tasks;Analysis tasks -> automatic classification;Analysis tasks -> pattern matching and detection;Musical features and properties","presentation":"Virtual","slackChannel":"p3-6-prima-la-musica","channelUrl":"https://ismir2026.slack.com/archives/C0C6278JFJQ","funFact":"Fun facts - \"Prima la Musica!\" (E. Cipollone and F. Cipollone), ISMIR 2026\n\n1. This is a father-and-son paper. Elvio is a composer and musicologist, with a PhD thesis on harmony in pop music; Francesco is a physicist, with a BSc thesis on subatomic particle decay, now doing a master's in Utrecht. The evaluation pipeline behind the published numbers was re-implemented from scratch, clean-room, by the physicist half of the team; his reproduction matched the released results.\n\n2. Most of the subsection titles in Section 3 are song titles: \"The Times They Are A-Changin'\", \"Unchained Melody\", \"I'd Like to Teach the World to Sing\", \"In the Air Tonight\", \"Pigs (Three Different Ones)\". The one exception is \"The Big Breath\": that phenomenon had no name, so we gave it one.\n\n3. Our affiliation, Centre Culturel BUQI, is a privately funded cultural centre in Le Mans, France. It opens officially in early 2027.","flags":[],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_56.html","displayId":"P3-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Structure, segmentation, and form","Analysis tasks","Automatic classification","Pattern matching and detection","Open Review"]},"72":{"uid":"72","title":"The Debussy Dataset: Evaluating Optical Music Recognition on Modern Staff Notation Manuscripts","authors":"Hugo Scheithauer; Thibault Clérice; Gonzalo Romero-García","abstract":"State-of-the-art (SOTA) Optical Music Recognition (OMR) models achieve low Symbol Error Rates (SER) on standard benchmarks, which consist almost exclusively of synthetic scores. Publicly available annotated datasets of handwritten historical manuscripts remain scarce, leaving the performance of current models on such sources understudied. We introduce the Debussy dataset, a corpus of 561 system-level and 101 full-page samples drawn from Claude Debussy's (1862–1918) autograph score manuscripts held at the French National Library, paired with MusicXML ground truth. The corpus spans 37 documents encompassing solo piano, double piano, and piano-voice layouts, and features handwriting-specific phenomena currently absent from synthetic benchmarks, such as erasures, idiosyncratic notations, and editorial annotations interleaved with musical content. Zero-shot evaluation of SOTA OMR models reveals near-total failure on this data with SER above 80%, exposing a large domain gap between training-distribution benchmarks and real-world historical sources. Fine-tuning experiments on our dataset reduce SER to 51–64%, confirming that even limited handwritten data can reduce this gap and offering a promising path toward more robust recognition of historical manuscripts.","primarySubject":"Evaluation, datasets, and reproducibility","secondarySubject":"Analysis tasks -> music transcription and annotation;Analysis tasks -> optical music recognition;Applications -> digital libraries and archives;Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"In-person","slackChannel":"p3-7-the-debussy-dataset","channelUrl":"https://ismir2026.slack.com/archives/C0C6278AT6Y","funFact":"Doing his PhD with classical music data, but only plays jazz guitar outside the office.","flags":["student"],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_72.html","displayId":"P3-07","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Analysis tasks","Music transcription and annotation","Optical music recognition","Applications","Digital libraries and archives","Evaluation methodology","Novel datasets and use cases","Open Review"]},"3":{"uid":"3","title":"Off-Manifold Robustness in Synthesizer Inversion with Joint Distribution Flow Matching","authors":"Ben Hayes","abstract":"Recent work on synthesizer inversion shows that generative models outperform deterministic approaches by explicitly modeling the ambiguity in mapping audio to parameters. Training such models, however, requires audio-parameter pairs, which are typically obtained by rendering sampled or preset parameters through the synthesizer itself. This creates a train-test mismatch that can degrade performance on off-manifold real-world recordings, for which ground-truth parameter annotations do not exist. To circumvent this obstacle, we propose to model the joint distribution of audio and parameters with a multi-modal continuous normalizing flow using independent noise schedules for each modality. This formulation allows us to train joint and conditional densities with paired synthesizer data, while unpaired real recordings can train the audio marginal alone, exposing the model to off-manifold signals without requiring parameter labels. Further, because the model learns to map from audio to parameters at all noise levels, we find that partially noising the audio reference at inference improves real-audio reconstruction, consistent with reducing sensitivity to distribution-specific detail while preserving coarse structure. Evaluating on Surge XT and Dexed, we find that modelling the joint distribution substantially improves both inversion of real-world and in-domain audio.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Applications -> music composition, performance, and production","presentation":"Virtual","slackChannel":"p3-8-off-manifold-robustness","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5S01PH","funFact":"","flags":[],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_3.html","displayId":"P3-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Applications","Music composition, performance, and production","Open Review"]},"112":{"uid":"112","title":"Analyzing and Reducing the Synthetic-to-Real Transfer Gap in Music Information Retrieval: The Task of Automatic Drum Transcription","authors":"Mickaël Zehren; Marco Alunno; Paolo Bientinesi","abstract":"Automatic drum transcription is a key task in Music Information Retrieval for extracting and analysing the rhythm of a music track. When solved with machine learning techniques, its performance is known to be limited by the size of the training data. A solution to increase the amount of data is to generate synthetic tracks from music scores rendered with virtual instruments. While this approach produces an arbitrarily large quantity of data, evidence shows that models trained on synthetic datasets do not transfer well to real-world tracks. To address this issue, we investigate the synthetic-to-real transfer gap and identify three strategies to make synthetic data more effective, namely, increasing the quality of the MIDI annotations, the presence of other instruments beyond drums, and the number of synthesizer configurations. For each one of these strategies, we measure how the performance of a model scales with varying amounts of data, and we apply scaling laws to determine at which point this performance stagnates. Ultimately, we conjecture that for training purposes synthetic data might be almost as effective as real-world audio.","primarySubject":"Evaluation, datasets, and reproducibility","secondarySubject":"Analysis tasks -> music transcription and annotation","presentation":"Virtual","slackChannel":"p3-9-analyzing-and-reducing","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7UCA2H","funFact":"Mickaël is also a drummer, he can play 16 notes a second on the bass drums.","flags":[],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_112.html","displayId":"P3-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Analysis tasks","Music transcription and annotation","Open Review"]},"120":{"uid":"120","title":"Steering Dense Music Retrieval With Open-Vocabulary Concept Discovery","authors":"Julien Guinot; Alain Riou; Elio Quinton; György Fazekas","abstract":"Controllable music retrieval lets users find music that is, for example, more ambient, less distorted, or without guitar while preserving the rest of an original seed query. Sparse autoencoders (SAEs) are a promising interface for this kind of concept-level control, but a key problem remains: given a free-form text concept, which sparse features should be edited? In shared multimodal embedding spaces, standard attribution methods often select neurons that match the concept’s wording but not the audio examples that express it. This leads to weak or unstable edits: relevant features are missed when concepts are distributed across neurons, while others are selected due to text alignment rather than audio-side structure.\n\nWe address this with a lightweight, training-free method that recovers a sparse set of audio features whose decoded representation reconstructs the target concept while remaining consistent with audio-space geometry. This reframes concept attribution as a sparse inversion problem rather than a text-side neuron-ranking heuristic. The method requires neither paired audio--text supervision nor SAE retraining.We evaluate this approach in steerable music retrieval and show that the recovered supports align more closely with concept-bearing audio examples and achieve a stronger trade-off between edit strength and preservation than alignment baselines, enabling more precise concept amplification and suppression with reduced drift on preservation metrics.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Applications -> music retrieval systems;MIR data fundamentals -> lyrics and other textual data;MIR data fundamentals -> multimodality","presentation":"Virtual","slackChannel":"p3-10-steering-dense-music","channelUrl":"https://ismir2026.slack.com/archives/C0C6279HZFW","funFact":"","flags":["student"],"session":"PS-03","posterUrl":"https://ismir2026program.ismir.net/poster_120.html","displayId":"P3-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Applications","Music retrieval systems","MIR data fundamentals","Lyrics and other textual data","Multimodality","Open Review"]},"35":{"uid":"35","title":"Multi-Task Multi-Frame Visual Piano Transcription","authors":"Yonghyun Kim; Hoyeol Sohn; Juhan Nam; Alexander Lerch","abstract":"Audio-based piano transcription performs well on onset, pitch, and velocity, but the sustain pedal lets sound persist long after key release, so audio systems predict pedal-extended offsets rather than physical key release. Yet existing Visual Piano Transcription (VPT) systems focus on onset detection from short video windows, offset accuracy lags onset by a wide margin, and note-level velocity has not been reported. To address these gaps, we present V2N (Video to Notes), the first complete VPT system: a shared temporal backbone feeds task-specific heads for onset, offset, key hold, and velocity, jointly trained with per-frame supervision rather than only at the window center. Ablations show that multi-task supervision enables offset and velocity prediction while improving onset accuracy; longer temporal context yields further improvements. V2N sets new state-of-the-art results on PianoVAM and R3.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology;MIR data fundamentals -> multimodality;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p4-1-multi-task-multi","channelUrl":"https://ismir2026.slack.com/archives/C0C6STFH4F2","funFact":"Yonghyun has taught both piano and AI programming at KAIST, two very different ways to put a keyboard to work!","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_35.html","displayId":"P4-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Evaluation, datasets, and reproducibility","Evaluation methodology","MIR data fundamentals","Multimodality","Musical features and properties","Representation learning","Open Review"]},"78":{"uid":"78","title":"H2H Music Improv: A Communication Model and Audio-Visual Dataset for Music Improvisation","authors":"Aleksandra Teng Ma; Anthony Cammarota; Jiayi Wang; Alexandria Smith; Cheng-Zhi Anna Huang; Jeffrey Albert; Alexander Lerch","abstract":"Current real-time AI improvisation systems lack the communication awareness human musicians rely on: rather than treating communication as a foundational algorithm design concern, most systems layer interaction strategies post-hoc onto generative algorithms through explicit controls and predefined modes. This gap persists in part because no formalized, machine-readable communication model with musicians exists. To address this, we study how expert musicians communicate in free (non-idiomatic) improvisation, unconstrained by prior discussion or agreement. Through a collaborative co-design process with expert improvisers, we derive a communication model that (1) captures how free improvisers negotiate musical ideas and enter stable musical spaces, and (2) is formalized as a machine-readable annotation scheme. We further present the H2H (Human-to-Human) Music Improvisation dataset: six hours of audio-visual expert duo improvisations with clean per-player stems and per-player annotations of both their own intentions and their perception of their partner's intentions. To our knowledge, this is the first such dataset for free improvisation. Together, the communication model and the dataset offer a new lens and resource for studying musician communication and may in future inform the design of AI musical partners that communicate by design.","primarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","secondarySubject":"Cognitive and user-centered MIR -> human-computer interaction;Cognitive and user-centered MIR -> user modeling;Evaluation, datasets, and reproducibility -> annotation protocols;MIR data fundamentals -> multimodality;Musical features and properties -> structure, segmentation, and form","presentation":"Virtual","slackChannel":"p4-2-h2h-music-improv","channelUrl":"https://ismir2026.slack.com/archives/C0C6278PS4U","funFact":"","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_78.html","displayId":"P4-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Novel datasets and use cases","Cognitive and user-centered MIR","Human-computer interaction","User modeling","Annotation protocols","MIR data fundamentals","Multimodality","Musical features and properties","Structure, segmentation, and form","Open Review"]},"80":{"uid":"80","title":"CODA: Cascaded Online Discontinuity-Aware Alignment for Real-Time Image-Based Score Following","authors":"Yining Yang; Ruogu Chen; Jie Han","abstract":"Real-time score following from sheet images remains challenging because the model must process streaming audio while resolving highly repetitive visual patterns under strict latency constraints. Recent image-based methods have attempted to use multi-resolution prediction by simultaneously predicting the positions of the active system, bar, and note. However, their predictions across these different levels of notation are independent, which makes the predictions unstable and introduces unnecessary extra search space for bar- and note-level predictions. Most existing methods also lack mechanisms to recover from score discontinuities, such as repeats, da capo (D.C.), or coda jumps. This paper proposes CODA, to the best of our knowledge, the first real-time score following system that addresses both gaps. CODA explicitly exploits the cascaded structure of music scores: it first selects the active system, then the active bar within it, and finally the active note within the selected bar. This enforces prediction consistency across resolutions. A silence-driven break mode enables recovery from arbitrary score discontinuities without requiring knowledge of the repeat structure. Evaluated on the Multimodal Sheet Music Dataset (MSMD) piano benchmarks, CODA achieves state-of-the-art tracking accuracy and discontinuity-recovery performance under real-time throughput. Code is available at https://anonymous.4open.science/r/CODA80.","primarySubject":"Analysis tasks -> alignment, synchronization, and score following","secondarySubject":"Applications -> music composition, performance, and production;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Generative Tasks -> real-time considerations;MIR data fundamentals -> multimodality;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p4-3-coda-cascaded-online","channelUrl":"https://ismir2026.slack.com/archives/C0C6278LG2Y","funFact":"CODA stands for Cascaded Online Discontinuity-Aware Alignment. Its name is a musical pun twice over.\nIn sheet music, “D.S. al Coda” sends performers back to the segno. They then play to the “To Coda” mark and jump to the coda. Our CODA tackles the challenge of following performers through repeats and jumps.\n“Coda” also means “tail” in Italian. That makes it a fitting name for a system designed to stay on the performer’s tail.","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_80.html","displayId":"P4-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Alignment, synchronization, and score following","Applications","Music composition, performance, and production","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Generative Tasks","Real-time considerations","MIR data fundamentals","Multimodality","Musical features and properties","Representation learning","Open Review"]},"146":{"uid":"146","title":"Improving Music Source Separation with Diffusion and Consistency Refinement","authors":"Tornike Karchkhadze; Mohammad Rasool Izadi; Shuo Zhang; Shlomo Dubnov","abstract":"In this work, we propose an approach to music source separation that uses a generative diffusion model as a last-stage refinement on top of a deterministic separator, progressively enhancing the separated sources through iterative denoising. While the diffusion refinement yields measurable quality gains, it requires iterative steps at inference, increasing computational cost. To speed up the inference process, we apply consistency distillation, reducing inference to a single step while maintaining quality; with two or more steps, the distilled model even surpasses the diffusion-based approach. Crucially, our method is architecture-agnostic: we demonstrate state-of-the-art results when applied to both a custom U-Net-based separator on Slakh2100 and the state-of-the-art BS-RoFormer model on MUSDB18, showing that the refinement generalizes across backbone architectures. Sound examples are available at: https://consistency-separation.github.io/","primarySubject":"Analysis tasks -> sound source separation","secondarySubject":"Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p4-4-improving-music-source","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7VKCJ1","funFact":"","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_146.html","displayId":"P4-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Sound source separation","Generative Tasks","Music and audio synthesis","Open Review"]},"218":{"uid":"218","title":"Do Music Generative Models Understand Musical Qualities? Automatic Music Evaluation with Model-Intrinsic Signals","authors":"Xiaosha Li; Chun Liu; Ziyu Wang","abstract":"Current music generative models can produce high-quality music, but does this ability imply that they “understand” the musical qualities of their outputs, and is that understanding aligned with human evaluation? Previous attempts to use the likelihood of a generative model to evaluate music, an approach commonly used in text, have proven unsuccessful, leading researchers to rely on standalone supervised music evaluation models. In this paper, we answer this question affirmatively: we show that a models intrinsic signals—derived from its hidden representations and predictions— are strongly correlated with human ratings. In particular, we study MusicGen and consider three types of features: (1) prediction loss, (2) prediction entropy, and (3) concepts extracted from the model using sparse autoencoders (SAE). Using these features, we train a lightweight prediction model to estimate subjective ratings. We evaluate these features both individually and in combination. We hypothesize that these signals parallel the listening process: the temporal and frequency-domain structure of loss and entropy reflect listeners’ expectation and surprise, while gradient directions in SAE latent space predict perceived quality. Experiments on five human-evaluation benchmarks spanning continuous ratings and pairwise preferences confirm this hypothesis, with SAE latents carrying most of the predictive signal.","primarySubject":"Evaluation, datasets, and reproducibility","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> evaluation metrics","presentation":"Virtual","slackChannel":"p4-5-do-music-generative","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA8GK2P","funFact":"We gave a music generator a second job: helping evaluate music. It keeps all its original weights; only a small rating model is trained.\n\nEven a music model can be “confidently wrong”: clips with more of these moments tended to get lower human ratings. Surprise when the model was uncertain showed the opposite pattern.\n\nAn AI captioner called two of our examples “pop, upbeat.” A music producer heard clear pop in one, but poor vocals and noisy percussion in the other. Hear the difference in our two-second demos. https://yoev.github.io/MEva/","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_218.html","displayId":"P4-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Evaluation methodology","Evaluation metrics","Open Review"]},"370":{"uid":"370","title":"Smorph: Playable Sound Morphing with Diffusion Models","authors":"Annie Chu; Hugo Flores García; Johannes Imort; Oriol Nieto; Bryan Pardo; Jordan Rudess; Prem Seetharaman; Justin Salamon","abstract":"Sound morphing, generating intermediate sounds that transition from one sonic identity to another, can be a powerful tool for musical sound design. Existing diffusion-based morphing approaches entangle temporal structure and timbral identity, offering no mechanism to hold one fixed while transforming the other. We present smorph, a training-free guidance framework that preserves how a sound behaves over time while transforming what the sound is, allowing users to morph, for instance from brass to strings at a fixed pitch. We demonstrate across three morphing modes: prompt-to-prompt, audio-to-prompt, and audio-to-audio. Evaluations across diverse datasets show that smorph effectively produces smooth morph trajectories while substantially improving temporal-structure and source preservation over baselines, albeit with more conservative target-ward transformation in some settings. In an exploratory case study, musicians found smorph trajectories to be expressive and playable, suggesting structural anchoring can serve as a productive constraint for instrumental interaction.","primarySubject":"Generative Tasks -> artistically-inspired generative tasks","secondarySubject":"Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p4-6-smorph-playable-sound","channelUrl":"https://ismir2026.slack.com/archives/C0C627K9HFW","funFact":"","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_370.html","displayId":"P4-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Artistically-inspired generative tasks","Music and audio synthesis","Open Review"]},"403":{"uid":"403","title":"Do Joint Language-Audio Embeddings Encode Perceptual Timbre Semantics?","authors":"Qixin Deng; Bryan Pardo; Thrasyvoulos N. Pappas","abstract":"Understanding and modeling the relationship between language and sound is essential for applications such as music information retrieval, text-guided music generation, and audio captioning. Central to these tasks are joint language-audio embedding spaces, which map textual descriptions and auditory content into a shared representation. Although multimodal embedding models such as MS-CLAP, LAION-CLAP, MuQ-MuLan, and OpenFLAM have shown strong performance in language-audio alignment, their correspondence to human perception of timbre, a multifaceted attribute encompassing qualities such as brightness, roughness, and warmth, remains under-explored. In this paper, we evaluate these joint language-audio embedding models in terms of their ability to capture perceptual timbre semantics. Across two complementary experiments, we find that LAION-CLAP shows relatively strong and consistent alignment with human-perceived timbre semantics across both instrumental sounds and descriptor-conditioned audio effects. At the same time, the overall strength of this alignment remains limited, suggesting that current joint language-audio embeddings capture perceptual timbre semantics only partially. We also observe that, overall, reverb-induced timbre semantics are more consistently encoded than equalization-induced timbre semantics.","primarySubject":"Evaluation, datasets, and reproducibility","secondarySubject":"Cognitive and user-centered MIR -> perception and cognition;Evaluation, datasets, and reproducibility -> evaluation methodology;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"Virtual","slackChannel":"p4-7-do-joint-language","channelUrl":"https://ismir2026.slack.com/archives/C0C60FEBZS5","funFact":"The first author met the second author during the ISMIR conference in 2024.\nThis work is inspired by Text2FX: Harnessing CLAP Embeddings for Text-Guided Audio Effects by Annie Chu","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_403.html","displayId":"P4-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Cognitive and user-centered MIR","Perception and cognition","Evaluation methodology","Musical features and properties","Timbre, instrumentation, and singing voice","Open Review"]},"409":{"uid":"409","title":"Assessing Factual Music Comprehension in Large Audio Language Models","authors":"Daniel Chenyu Lin; Michael Freeman; John Thickstun","abstract":"Large audio language models (LALMs) leverage multimodal representations to generate open-ended answers to natural language queries about audio. In this paper, we (1) provide empirical evidence that assessment of LALMs using the popular MusicQA dataset fails to measure whether a model's responses about music are factually correct, and (2) develop a new protocol for assessing the music comprehension capabilities of LALMs. Specifically, we propose an evaluation protocol that prompts a LALM for factually verifiable information, and parses its open-ended response into a structured format that can be objectively assessed using Precision, Recall, and F1 scores. Using this protocol, we define a benchmark consisting of six factual information retrieval tasks defined on three diverse datasets: MusicNet, the Free Music Archive, and OverClocked ReMix. We benchmark nine recent LALMs, including frontier models like Gemini and the latest open models like Music Flamingo, and release the suite of evaluation scripts at https://github.com/DCL2004/LALM-Eval to facilitate benchmarking of new LALMs.","primarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology","secondarySubject":"","presentation":"Virtual","slackChannel":"p4-8-assessing-factual-music","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6875AB","funFact":"","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_409.html","displayId":"P4-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Evaluation methodology","Open Review"]},"418":{"uid":"418","title":"Spacing Out: On the Reliability of Binaural Music Source Separation Metrics","authors":"Richa Namballa; Magdalena Fuentes","abstract":"Despite the rising popularity of immersive audio, binaural music remains underexplored in music information retrieval (MIR), particularly regarding the task of music source separation (MSS). While existing stereo MSS models can process binaural audio, they often degrade the spatial quality of the separated stems and undermine listener immersion. Through a perceptual study comparing binaural and stereo MSS outputs, we evaluate how well objective spatial distortion metrics correlate with human perception. Our findings reveal varied agreement between these metrics and human judgment, highlighting a lack of reliability when used to evaluate binaural music tasks. Specifically, we find that Interaural Time Difference (ITD) estimation is highly sensitive to noise and separation artifacts. In evaluating two alternative ITD estimation methods, we uncover a critical trade-off between robustness and accuracy, particularly for narrow-band instruments like bass. These results underscore the need for accurate, interpretable spatial metrics designed for binaural music to develop models that preserve source localization and listener immersion.","primarySubject":"Analysis tasks -> sound source separation","secondarySubject":"Cognitive and user-centered MIR -> user-centered evaluation;Evaluation, datasets, and reproducibility -> evaluation metrics;MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p4-9-spacing-out-on","channelUrl":"https://ismir2026.slack.com/archives/C0C5YC22YRY","funFact":"This is a follow-up paper to \"Do Music Source Separation Models Preserve Spatial Information in Binaural Audio?\" from ISMIR 2025.\n\nRicha would like to shout-out her fellow chairs of the Late-Breaking/Demo Committee, Liwei and Lucas, for their hard work!\n\nRicha fosters cats and dogs from the local city shelter. Her current foster cat, Cooper (f.k.a. Colossus), is 21 lbs (9.5 kg).","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_418.html","displayId":"P4-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Sound source separation","Cognitive and user-centered MIR","User-centered evaluation","Evaluation, datasets, and reproducibility","Evaluation metrics","MIR data fundamentals","Music signal processing","Open Review"]},"444":{"uid":"444","title":"Local Multimodal Music Alignment from Global Supervision","authors":"Irmak Bukey; Zachary Novack; Jongmin Jung; Dasaem Jeong; Chris Donahue","abstract":"Understanding music requires understanding localized relationships across data modalities, e.g., how time in performance audio maps onto position in a score image. Yet supervision for such local correspondences is difficult to obtain—in practice, we often only have access to coarser global supervision like paired segments of audio and images. To address this gap, we propose FuSiLi (Fused Sinkhorn-Localized Similarity), a similarity score for multimodal contrastive learning operating directly on local image patch and audio frame features via Sinkhorn-based soft alignment. We show that FuSiLi (i) effectively learns local relationships, (ii) requires only global supervision, and (iii) retains the global alignment capabilities of conventional contrastive approaches. We fine-tune pretrained CLIP and CLAP encoders on pairs of raw sheet music images and audio using a hybrid contrastive objective combining FuSiLi with conventional global similarity. We evaluate on cross-modal retrieval and frame-level alignment tasks against a range of global and local baselines, showing that our approach outperforms them on local alignment while remaining competitive on retrieval.","primarySubject":"MIR data fundamentals -> multimodality","secondarySubject":"Analysis tasks -> alignment, synchronization, and score following;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p4-10-local-multimodal-music","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6G4DPZ","funFact":"","flags":["student"],"session":"PS-04","posterUrl":"https://ismir2026program.ismir.net/poster_444.html","displayId":"P4-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Multimodality","Analysis tasks","Alignment, synchronization, and score following","Musical features and properties","Representation learning","Open Review"]},"371":{"uid":"371","title":"Does Notation Matter? MusicXML Tokenisation for Symbolic Music Generation and Understanding","authors":"Andrea Poltronieri; Matteo Spanio; Simone Chieppa; Xavier Serra; Martín Rocamora","abstract":"Symbolic music representations encode musical content at varying levels of completeness, from performance-oriented formats such as MIDI to fully notated scores such as MusicXML. Yet the impact of this representational richness on deep learning models remains poorly understood. While tokenisation has been extensively studied for MIDI, score-based formats have received considerably less attention, and MusicXML in particular remains largely unexplored despite being the de-facto standard for score interchange. In this paper, we introduce the first general-purpose tokenisation framework for MusicXML, and conduct a systematic empirical study to determine whether notational richness confers a measurable advantage for deep learning on symbolic music. We train 26 autoregressive and masked language models, comparing MusicXML tokenisation against MIDI and ABC-based representations, probing the effect of different tokenisation strategies within MusicXML, and testing whether music-grammar-aware tokenisation outperforms treating scores as plain text. We evaluate music understanding tasks using linear probing across composer classification, difficulty estimation, and accompaniment suggestion, and generative tasks using objective statistical metrics and a human pairwise preference study. Our results show that MusicXML-tokenised models outperform MIDI and text-based counterparts on understanding and generative tasks.","primarySubject":"MIR data fundamentals -> symbolic music processing","secondarySubject":"Musical features and properties -> representation learning;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p5-1-does-notation-matter","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCNA69G","funFact":"The models in this paper were trained on four different clusters in four different countries.","flags":[],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_371.html","displayId":"P5-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Symbolic music processing","Musical features and properties","Representation learning","Representations of music","Open Review"]},"261":{"uid":"261","title":"Music Genre Classification Using Audio and Automatically Transcribed Lyrics","authors":"Jun-You Wang; Hsuan Yu Chen","abstract":"Recent audio-text multimodal systems have achieved improved performance in music genre classification over single-modal systems. However, a crucial limitation is that multimodal systems require music data with multiple modalities for training and inference, which is not always available. In this work, we propose using a multilingual lyrics transcription model to transcribe lyrics from input music audio, and utilize these automatically generated transcriptions along with the original audio for music genre classification. Experiment results show that the proposed method consistently improves the performance of multiple datasets over its single-modal counterpart by up to 4.4% PR-AUC score, without the need for additional lyrical annotation as input. Further analysis reveals that the proposed method is particularly effective when the input audio contains a full song instead of a short excerpt. To our knowledge, this is the first study to utilize automatically transcribed lyrics to enhance music classification performance, opening up a new possibility for future music genre classification research.","primarySubject":"Musical features and properties -> musical style and genre","secondarySubject":"Analysis tasks -> automatic classification;MIR data fundamentals -> lyrics and other textual data;MIR data fundamentals -> multimodality","presentation":"In-person","slackChannel":"p5-2-music-genre-classification","channelUrl":"https://ismir2026.slack.com/archives/C0C6STLRLTA","funFact":"","flags":[],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_261.html","displayId":"P5-02","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Musical style and genre","Analysis tasks","Automatic classification","MIR data fundamentals","Lyrics and other textual data","Multimodality","Open Review"]},"267":{"uid":"267","title":"BARS: Beat-Adaptive Rap Synthesis via Flow Matching","authors":"Taemin Cho; Hector Martel; Enyan Koh","abstract":"We present BARS (Beat-Adaptive Rap Synthesis), a non-autoregressive rapping voice generation model based on Flow Matching, conditioned on a beat track (instrumental backing), lyrics, and a short reference voice for zero-shot timbre control. Unlike prior autoregressive rap synthesis, where rap pace and output duration are strictly determined by the beat, BARS enables independent control over the rap tempo, allowing flexible-length generation without explicit duration prediction or phoneme-level alignment.\nWe encode lyrics using a pretrained ByT5 encoder with a lightweight Transformer adapter trained jointly with the model, without relying on speech-pretrained text encoders, enabling natural handling of non-standard orthography (e.g., slang, abbreviations) and numerals common in rap lyrics.\nTo address alignment instabilities, such as token skipping or repetition, under strong beat conditioning, we propose Asymmetric Scaled RoPE for cross-attention, which introduces a positional structure to facilitate sequential lyric-audio alignment learning. Furthermore, the model supports partial re-generation via span masking, enabling lyric-level editing while preserving the surrounding vocal style. \nExperiments demonstrate that the proposed model achieves high lyric fidelity, natural rap flow that is rhythmically coherent with the beat, and robust tempo control across a wide range of rapping styles and speeds.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Cognitive and user-centered MIR -> user-centered evaluation;Creativity -> human-AI co-creativity;Creativity -> tools for artists;Generative Tasks -> artistically-inspired generative tasks;Musical features and properties -> rhythm, beat, tempo","presentation":"In-person","slackChannel":"p5-3-bars-beat-adaptive","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5XULS3","funFact":"Taemin Cho’s talk ends with an AI-generated rap in his own voice. Some say it’s easier to understand than the talk itself, thanks to its clearer pronunciation.","flags":["award"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_267.html","displayId":"P5-03","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Cognitive and user-centered MIR","User-centered evaluation","Creativity","Human-AI co-creativity","Tools for artists","Artistically-inspired generative tasks","Musical features and properties","Rhythm, beat, tempo","Awards Nominee"]},"271":{"uid":"271","title":"Musical Emotion under Cochlear Implant Listening: Human Perception and Foundation Model Representations of Human-Composed and AI-Generated Music","authors":"Taein Song; Hyunjae Kim; Kyung Myun Lee","abstract":"Perceiving diverse emotions in music, driven by multiple acoustic cues, plays a central role in human life. However, for Deaf and Hard-of-Hearing populations, including cochlear implant (CI) users, degraded acoustic cues can limit access to musical emotion. Motivated by recent advances in generative AI, we investigate whether such systems can support more accessible music experiences. We conducted a behavioral experiment in which CI and normal-hearing (NH) groups rated emotions of human-composed and AI-generated music using emotion-category ratings and a valence–arousal (VA) scale. We further analyzed foundation model representations to examine how musical emotions are encoded and how these representations align with human responses. Behaviorally, emotion categories were less distinguishable in the CI group than in the NH group for human-composed music, and this effect was more pronounced for AI-generated music. Model analyses showed a similar but model-dependent pattern: several representations exhibited reduced emotion separability under CI-simulation, while VA prediction suggested that CI-simulated embeddings better captured CI users' valence ratings than original-audio embeddings. Together, these findings show that musical emotion is reorganized under CI-related listening conditions and that this reduction in perceptual differentiation is pronounced for AI-generated music, underscoring the need for accessibility-aware evaluation of music AI across diverse auditory populations.","primarySubject":"Musical features and properties -> musical affect, emotion and mood","secondarySubject":"Applications -> music and health, well-being and therapy;Cognitive and user-centered MIR -> perception and cognition;Cognitive and user-centered MIR -> user behavioral and physiological responses;Cognitive and user-centered MIR -> user-centered evaluation;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p5-4-musical-emotion-under","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJ78N06","funFact":"Taein is a big fan of \"Jordy\", one of Kakao's characters - a cute mint-colored dinosaur. She has tons of Jordy merchandise, so her home basically looks like a giant mint-colored world!","flags":["student"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_271.html","displayId":"P5-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Musical affect, emotion and mood","Applications","Music and health, well-being and therapy","Cognitive and user-centered MIR","Perception and cognition","User behavioral and physiological responses","User-centered evaluation","Representations of music","Open Review"]},"251":{"uid":"251","title":"Do Music Foundation Models Embed Pitch in Helical Structure?","authors":"Hayato Yagi; Shinnosuke Takamichi; Rin Sato; Keitaro Tanaka; Shigeo Morishima","abstract":"This study analyzes the intermediate representations of music foundation models (MFMs) and reports the geometric structures used to represent pitch information. By inputting isolated musical notes into trained MFMs and analyzing their principal components, we reveal that the representations form a helical structure reflecting the octave periodicity of pitch. Furthermore, we show that the clarity and geometry of this helical structure vary not only across models but also with the acoustic properties of the input signals. Our analysis provides a novel approach for clarifying the internal mechanisms of MFMs.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology;MIR data fundamentals -> music signal processing;Musical features and properties -> representations of music;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"Virtual","slackChannel":"p5-5-do-music-foundation","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7VABHB","funFact":"","flags":["student"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_251.html","displayId":"P5-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Evaluation, datasets, and reproducibility","Evaluation methodology","MIR data fundamentals","Music signal processing","Representations of music","Timbre, instrumentation, and singing voice","Open Review"]},"318":{"uid":"318","title":"Beyond Dry References: Learning Relative Audio Effects Representations via Contrastive Distance Learning","authors":"Xinlu Liu; Huibin Lin; Weixing Wei; Zhenhai Yan","abstract":"Audio effects (Fx) representation learning plays a key role in intelligent music production, including automatic mixing and Fx style transfer. Existing methods typically rely on dry or nearly dry references for effect modeling, yet truly unprocessed audio is rarely available in practice, as real recordings inevitably reflect the microphone, room acoustics, and preceding signal processing. Instead of pursuing absolute effect encodings, we argue that the relative effect distance between audio signals is more meaningful for real-world music production. Motivated by this, we propose RelFx, a contrastive learning framework that learns relative effect transformations from general audio collections without requiring dry references during representation training. Our approach uses a dual-branch Siamese encoder equipped with cross-attention and differential gating fusion to infer the shared effect transformation from a reference clip and an effect-processed, content-related clip. We further propose an antisymmetric fusion variant for bidirectional effect encoding, such that swapping the input order directly produces a nearly sign-reversed embedding, a property not explored in earlier work. Moreover, our dry-reference-free formulation eliminates the reliance on dry multitrack datasets and enables training on effect-bearing audio. Experiments on Fx style transfer demonstrate state-of-the-art performance under the standard Fx-Encoder++ MUSDB18 evaluation protocol, consistently outperforming existing approaches across all four instrument categories.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks -> similarity metrics;Applications -> music composition, performance, and production;Musical features and properties -> representations of music","presentation":"In-person","slackChannel":"p5-6-beyond-dry-references","channelUrl":"https://ismir2026.slack.com/archives/C0C5SABMLQ3","funFact":"","flags":[],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_318.html","displayId":"P5-06","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Similarity metrics","Applications","Music composition, performance, and production","Representations of music","Open Review"]},"321":{"uid":"321","title":"Stem-Specialized Multi-Gate Mixture-of-Experts for Joint Music Structure Analysis","authors":"Aneeka Azmat; Li Su; ChengHsin Hsu","abstract":"Jointly solving boundary detection and functional labeling in Music Structure Analysis (MSA) remains challenging because the two tasks draw on fundamentally different acoustic cues: boundaries manifest through abrupt local changes, while functional labels reflect the sustained semantic character of entire sections. Existing models address both tasks with a single shared representation, ignoring this asymmetry. To address this issue, we propose a Multi-gate Mixture-of-Experts (MMoE) framework that assigns each task its own gating network to dynamically weight a pool of multiple stem-specialized expert encoders. Each expert extends a dual-attention MSA architecture by incorporating a stem encoder into the cross-attention fusion to model multifaceted similarity in music. Experiments using Beatles, SALAMI, and RWC-Pop datasets demonstrate that the proposed MMoE framework outperforms all baselines on both boundary detection and functional labeling metrics, confirming that task-specific dynamic gating over stem-specialized experts provides a more effective joint MSA representation than a single shared model.","primarySubject":"Analysis tasks","secondarySubject":"Analysis tasks -> sound source separation;Evaluation, datasets, and reproducibility -> evaluation metrics","presentation":"In-person","slackChannel":"p5-7-stem-specialized-multi","channelUrl":"https://ismir2026.slack.com/archives/C0C60F96HTK","funFact":"1: Same four instrument experts, two gates and each task listens to a different part of the band.\n2: To learn where the MMoE idea comes from, see Ma et al., Multi-Gate Mixture-of-Experts, KDD 2018, originally built for recommendations, now finding song structure.","flags":["student"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_321.html","displayId":"P5-07","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Analysis tasks","Sound source separation","Evaluation, datasets, and reproducibility","Evaluation metrics","Open Review"]},"424":{"uid":"424","title":"Live Music Denoising with Authentic Live Data and Track-Editing-Guided Semi-Supervised Learning","authors":"Penghao He; Xiankexin Luo; Wei Li","abstract":"Live music denoising for concerts, live houses, and other music performances has growing demand across music production, broadcasting, archiving, and audience-facing services, yet research in this area remains scarce and realistic data for training and benchmarking are lacking. To address these gaps, we collected clean music segments together with pure noise excerpts from real live performances. With the collected noise data, we produce noise tracks at multiple intensity levels to create controlled test conditions and benchmark a range of deep learning models. We then compared models trained with our live dataset against models trained only on more readily available but less representative data, such as filtered public noise samples and studio recordings, quantifying their performance differences on eliminating genuine live noise. To mitigate the practical difficulty of obtaining labeled live data, we introduce a semi-supervised training strategy that leverages additional unlabeled noisy live recordings and modifies pseudo-label tracks through targeted editing informed by class priors, yielding measurable gains in denoising quality. Experimental results demonstrate that models trained with authentic live material and the proposed semi-supervised refinement generalize better to real live recordings with different noise levels, underscoring the value of realistic data and tailored learning strategies for live music denoising.","primarySubject":"MIR data fundamentals -> music signal processing","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"Virtual","slackChannel":"p5-8-live-music-denoising","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCW81EE","funFact":"","flags":["student"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_424.html","displayId":"P5-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Music signal processing","Evaluation, datasets, and reproducibility","Novel datasets and use cases"]},"459":{"uid":"459","title":"Digital Existence as Ontological Intervention: Backstage Networks in Postwar Taiwanese Popular Music","authors":"Ming Cheng; Yung-Chuan Chang; Chien-Chang Yang; Li Su","abstract":"The historical narrative of popular music privileges front-stage stars, and music information retrieval (MIR) has largely reproduced this bias. Backstage laborers—arrangers, session musicians, engineers—remain absent from the field's datasets and research questions. Responding to Born's (2020) call for a critical MIR, we argue that encoding historically marginalized music workers into computational databases constitutes an ontological intervention that grants them digital existence in an era where absence from metadata means erasure from history. We introduce the Private Archive of Scanned Music Artifacts (PASMA), a dataset reconstructed by an LLM pipeline from 1,122 Taiwanese vinyl liner notes spanning 1957–1991, and combine Social Network Analysis with MERT embeddings on vocal-separated stems. The collaboration network transformed sharply around 1978, when multitrack recording shifted it from a fragmented, high-modularity ecology to a centralized giant component bridged by elite intermediaries. MERT embeddings indicate that the arrangement traditions defined by these network communities occupy distinct regions in acoustic space. These findings demonstrate that MIR's computational tools, redirected toward the backstage, can decode historical power structures and restore the visibility of the laborers who built the soundscapes of postwar Taiwan.","primarySubject":"Computational musicology","secondarySubject":"Computational musicology -> computational ethnomusicology;Computational musicology -> systematic musicology;Responsibility and Ethics in MIR -> philosophical and methodological foundations","presentation":"In-person","slackChannel":"p5-9-digital-existence-as","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJRS0LE","funFact":"NA","flags":["award"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_459.html","displayId":"P5-09","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Computational musicology","Computational ethnomusicology","Systematic musicology","Responsibility and Ethics in MIR","Philosophical and methodological foundations","Awards Nominee","Open Review"]},"485":{"uid":"485","title":"Frame-Level Pansori Mode Classification with Complementary Audio Representations","authors":"Sangheon Park; Seonguk Ju; Suin Chung; Danbinaerin Han; Dasaem Jeong","abstract":"Pansori is a traditional Korean vocal genre whose mode system (jo) is defined not by scale alone but by the entanglement of pitch collection, microtonal ornament (sigimsae), and vocal timbre. In this study, we introduce a 46-hour frame-level pansori mode annotation, expert-labeled across all five canonical batang, and evaluate four complementary input representations (mel spectrogram, F0 contour, MIDI piano roll, and a multi-cultural SSL encoder) under two split strategies designed to detect shortcut learning. Across the three well-represented modes, performance degrades by only 1.7--3.6 points of F1 when entire works are held out, indicating that the models learn mode-relevant features rather than memorizing repertoire. Per-class results further show that source separation removes the percussion cue on which changjo depends, and that generic multi-cultural pre-training fails specifically on the Ujo--Gyemyeonjo distinction. Qualitative analysis of cross-modal disagreement recovers musicologically documented phenomena and agrees with published score-based analyses of modern changjak pansori.","primarySubject":"Computational musicology -> computational ethnomusicology","secondarySubject":"Computational musicology -> computational music theory and musicology;Computational musicology -> digital musicology","presentation":"Virtual","slackChannel":"p5-10-frame-level-pansori","channelUrl":"https://ismir2026.slack.com/archives/C0C5SAULERH","funFact":"Sangheon Park made another presentation yesterday session 2. This is his last official oral presentation for this year's ISMIR!","flags":["student"],"session":"PS-05","posterUrl":"https://ismir2026program.ismir.net/poster_485.html","displayId":"P5-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Computational musicology","Computational ethnomusicology","Computational music theory and musicology","Digital musicology","Open Review"]},"164":{"uid":"164","title":"On the use of Pre-trained Audio Encoders for Audio-to-Score Music Transcription","authors":"Antonio Hidalgo-Centeno; Carlos Penarrubia; Jose J. Valero-Mas","abstract":"Audio-to-Score (A2S) transcription, a recent paradigm within Automatic Music Transcription (AMT), converts acoustic music recordings into score-level representations in a single step. These approaches primarily rely on end-to-end deep neural architectures, which avoid error propagation in multi-stage score-oriented AMT systems. However, they require large collections of music recordings paired with annotated scores, which are severely limited due to the substantial manual effort and expert knowledge required for their creation. Drawing inspiration from related domains, this work investigates the use of pre-trained audio encoders to mitigate this limitation and improve transcription accuracy. Specifically, we evaluate six encoders from different domains—including music, speech, and general audio processing—and assess their capabilities for music transcription. Experiments with a state-of-the-art A2S framework under different adaptation strategies on polyphonic music show that pre-training improves performance, achieving recognition error rates of below 3%. These findings suggest that pre-trained audio models are a promising direction for advancing A2S transcription and reducing dependence on large annotated datasets.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Evaluation, datasets, and reproducibility -> reproducibility;MIR data fundamentals -> music signal processing;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p6-1-on-the-use","channelUrl":"https://ismir2026.slack.com/archives/C0C5YC615U2","funFact":"This is the first participation of the first Author at ISMIR, of (we hope!) more to come.\nWe are also working on research on Multimodal Representations where music notation plays a major role.","flags":["student"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_164.html","displayId":"P6-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Evaluation, datasets, and reproducibility","Reproducibility","MIR data fundamentals","Music signal processing","Musical features and properties","Representation learning","Open Review"]},"184":{"uid":"184","title":"Integrating Contextual Embeddings into Evaluation of Expressive MIDI Piano Performances","authors":"Dmitrii Gavrilev; Ilya Borovik; Vladimir Viro","abstract":"Objective evaluation of expressive MIDI piano performances typically relies on attribute statistics such as timing, velocity, and duration of individual notes. However, these methods often disregard dependencies between notes, which poses a potential limitation in assessing the similarity between two sets of performances. In generative applications, the wide variety of expressive attributes makes it difficult to aggregate them into a single scalar metric for model selection. In this work, we reexamine attribute-scoped metrics and explore the perceptual properties of contextual embeddings from self-supervised symbolic music models, Aria and CLaMP3. Results from our listening study indicate that these models can be used as perceptual proxies, showing agreement with per-sample human ratings on par with traditional metrics. To measure conditional distributional similarity, we adapt Kernel Audio Distance to the symbolic music domain. Unlike Pearson correlation and reconstruction error, kernel-based methods on contextual embeddings do not require note alignment and are sensitive to contextual perturbations. To facilitate reproducibility, we release Pereval, an open-source library that integrates performance evaluation utilities, including both attribute-scoped and deep feature metrics.","primarySubject":"Evaluation, datasets, and reproducibility -> evaluation metrics","secondarySubject":"Analysis tasks -> similarity metrics;Generative Tasks -> evaluation metrics;MIR data fundamentals -> symbolic music processing;Musical features and properties -> expression and performative aspects of music","presentation":"Virtual","slackChannel":"p6-2-integrating-contextual-embeddings","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5SF97H","funFact":"","flags":["student"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_184.html","displayId":"P6-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Evaluation metrics","Analysis tasks","Similarity metrics","Generative Tasks","MIR data fundamentals","Symbolic music processing","Musical features and properties","Expression and performative aspects of music","Open Review"]},"187":{"uid":"187","title":"Beat This! Live - Online Beat and Downbeat Tracking by Predicting the Future","authors":"Bánk Juhász; Gerhard Widmer; Jan Schlüter","abstract":"Beat and downbeat tracking have seen multiple improvements through recent deep learning approaches, but most target an offline setting, precluding their use in real-time scenarios. We adapt a recent offline beat tracking model, Beat This! [1], for online, low-latency operation and propose Beat This! Live, a causal variant designed for real-time beat and downbeat tracking. We present the architectural modifications required to obtain a causal system suitable for streaming inference and measure their impact on accuracy. By training our model with temporally shifted labels, it produces predictions slightly ahead of the current frame, which not only reduces latency, but enables much more effective peak-picking. Evaluation on the GTZAN dataset shows the F1-score improves from 80.6 to 86.7 for beats and 56.5 to 74.1 for downbeats compared to BeatNet+ [2], the current state of the art in online beat tracking.","primarySubject":"Musical features and properties -> rhythm, beat, tempo","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation metrics;MIR data fundamentals -> music signal processing","presentation":"In-person","slackChannel":"p6-3-beat-this-live","channelUrl":"https://ismir2026.slack.com/archives/C0C60F1GB41","funFact":"","flags":["student"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_187.html","displayId":"P6-03","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Rhythm, beat, tempo","Evaluation, datasets, and reproducibility","Evaluation metrics","MIR data fundamentals","Music signal processing","Open Review"]},"188":{"uid":"188","title":"Latent Trace Guidance for Generative Control Without Disentangled Data in Symbolic Melodic Harmonization","authors":"Maximos Kaliakatsos-Papakostas; Dimos Makris; Konstantinos-Theodoros Tsamis; Konstantinos Soiledis; Vassilis Katsouros","abstract":"Controlling generative models without access to paired or disentangled data remains a fundamental challenge, particularly in domains where attributes such as style are inherently entangled with content. We introduce latent trace guidance, a general framework for controllable generation that operates by aligning a model’s internal representations with an external guidance signal expressed in a compatible latent space. Unlike conventional conditioning or activation steering, the proposed approach does not require explicit supervision or attribute labels, and allows guidance signals to originate from arbitrary sources, provided they can be mapped into the model’s representation space via an adapter. To isolate the core mechanism, we study a pure-form instantiation in the context of symbolic melodic harmonization, where guidance is derived directly from the model’s own latent representations and re-applied during generation. This setting enables a controlled comparison with activation steering across architectures and cross-style scenarios. Experimental results show that latent trace guidance consistently improves alignment with target stylistic characteristics and, in several cases, outperforms activation steering, particularly when counteracting strong model biases. These findings highlight latent-space alignment as a promising direction for controllable generation beyond explicitly supervised settings.","primarySubject":"Applications -> music composition, performance, and production","secondarySubject":"Creativity -> creative practice involving MIR or generative technology;MIR data fundamentals -> symbolic music processing;Musical features and properties -> harmony, chords and tonality","presentation":"Virtual","slackChannel":"p6-4-latent-trace-guidance","channelUrl":"https://ismir2026.slack.com/archives/C0C60F1D525","funFact":"","flags":[],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_188.html","displayId":"P6-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications","Music composition, performance, and production","Creativity","Creative practice involving MIR or generative technology","MIR data fundamentals","Symbolic music processing","Musical features and properties","Harmony, chords and tonality","Open Review"]},"192":{"uid":"192","title":"LiveBand: Live Accompaniment Generation in the Audio Domain","authors":"Marco Pasini; Javier Nistal; Mathias R. Bjare; Stefan Lattner; György Fazekas","abstract":"We present LiveBand, a real-time system that generates high-fidelity music accompaniments to live audio input, respecting strict causal constraints. Our method trains a causal transformer generator in the continuous latent space of a pre-trained causal audio autoencoder, using adversarial sequence-level supervision from a discriminator. At each timestep, the generator receives only the causally available mix context and Gaussian noise, and predicts accompaniment latents without access to future mix frames or ground-truth target latents. Training is performed in a single parallel forward pass under causal masking, while streaming inference proceeds autoregressively with a rolling attention state. The model’s training and inference computations are matched by design, eliminating teacher forcing and the associated exposure bias. On a multi-instrument music accompaniment benchmark, LiveBand improves over prior work on objective measures of audio quality, beat alignment, and mix adherence, while enabling real-time streaming generation without lookahead into the future on consumer hardware.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Generative Tasks -> real-time considerations","presentation":"Virtual","slackChannel":"p6-5-liveband-live-accompaniment","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA817D1","funFact":"","flags":["student","award"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_192.html","displayId":"P6-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Real-time considerations","Awards Nominee","Open Review"]},"279":{"uid":"279","title":"Soft Late-Fusion Alignment and Pattern-Based Performance Analysis of Early 20th-Century Solo Flute Music","authors":"Patrice Thibaud; Louis Bigo; Mathieu Giraud; Yann Teytaut","abstract":"This paper investigates expressive timing in solo flute performances of early 20th-century repertoire through curated data, audio-to-audio alignment, and pattern-level analysis. We introduce a corpus of 23 performances each of Debussy’s Syrinx and Varèse’s Density 21.5, enabling the study of timing in flexible metronomic contexts. A subset of recordings is manually annotated at the onset level and used as a reference to propagate annotations to the remaining performances via a multi-reference audio-to-audio alignment framework. We extend existing late-fusion strategies to onset-level transfer and introduce a soft weighting scheme based on triple-alignment consistency, replacing hard reference selection with a continuous reliability measure. This improves the stability of onset estimates across heterogeneous interpretations. We further propose a pattern-level framework for expressive timing that extends beyond beat-based representations, enabling systematic comparison of duration and timing fluctuations in both cell- and phrase-level patterns. This allows us to investigate whether performers realise these patterns consistently or variably across different levels of musical organisation. Overall, the proposed pipeline supports the alignment, propagation, and analysis of expressive timing in repertoires with flexible metronomic structure.","primarySubject":"Musical features and properties -> expression and performative aspects of music","secondarySubject":"Analysis tasks -> alignment, synchronization, and score following;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Evaluation, datasets, and reproducibility -> reproducibility;Musical features and properties -> melody and motives","presentation":"Virtual","slackChannel":"p6-6-soft-late-fusion","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCAFUV8","funFact":"🪈 In Greek mythology, Pan's flute was created from the reeds of the nymph Syrinx, who transformed to escape the god Pan.\n\n🎶 Our Corpus is available on Dezrann. Don't know this wonderful platform ? Let's take a look at https://www.dezrann.net/explore/early20thsoloflute\n\n⚖️ Density 21.5 refers to the density of a 90/10 platinum–iridium alloy used for Georges Barrère's flute.","flags":["student"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_279.html","displayId":"P6-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Expression and performative aspects of music","Analysis tasks","Alignment, synchronization, and score following","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Reproducibility","Melody and motives","Open Review"]},"196":{"uid":"196","title":"Playability-Aware Audio-to-Tablature Guitar Transcription via Diffusion Models","authors":"Riccardo Simionato; Louis Bigo","abstract":"Guitar tablature transcription requires not only accurate pitch detection but also assigning each note to a specific string-fret position, as the same pitch can be played at multiple fretboard positions. Existing approaches treat this as a standard classification problem, ignoring the musical and physical constraints that govern playable fingering sequences. We propose Noise2Fret, a diffusion model for audio-to-tablature transcription that generates tablature through a continuous latent representation of discrete fret and string targets, conditioned on spectral and audio features. To bridge the gap between pitch accuracy and physical playability, we introduce five auxiliary losses encoding Pitch-Class Distance, Positional Distance, Circle-of-Fifths Distance, String Similarity, and Hand-Span Feasibility directly into the training objective. Experiments on GuitarSet and GOAT datasets demonstrate that the model outperforms baselines while remaining computationally more efficient, and that the auxiliary losses yield consistent gains over the standard training objective.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Analysis tasks -> automatic classification;MIR data fundamentals -> symbolic music processing","presentation":"Virtual","slackChannel":"p6-7-playability-aware-audio","channelUrl":"https://ismir2026.slack.com/archives/C0C627B6SDA","funFact":"Working on Noise2Fret motivated Riccardo to finally pick up the guitar and start learning to play it himself","flags":[],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_196.html","displayId":"P6-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Automatic classification","MIR data fundamentals","Symbolic music processing","Open Review"]},"198":{"uid":"198","title":"OudTabs: An Interactive Microtonal Tablature Platform for Cross-Tradition Oud Repertoires","authors":"Anıl Tuncel","abstract":"This paper presents OudTabs, a microtonal oud tablature representation, a cross-tradition symbolic corpus, and an interactive software platform for makam-family and related modal repertoires. Its underlying representation aims to diversify music research beyond Western staff notation by grounding itself in the oud’s fingerboard, preserving each tradition’s native pitch system, modal vocabulary, and rhythmic categories. Every note is encoded as string, position, and duration on the instrument, with pitch expressed as exact frequency-derived fractional positions in cents. The same cent-level values drive both the visual display and the audio, so learners and researchers hear each microtone at its target pitch while simultaneously seeing it as a concrete location on the oud. This cent-level resolution lets the OudTabs representation capture microtonal intonation more finely than the notations commonly used in oud teaching. Built on this foundation, we release the OudTabs Corpus, a cross-tradition tablature corpus of 3,672 pieces and 1,957,167 note events unified from heterogeneous public symbolic corpora, made playable, searchable, and audible through an interactive playback and browsing system. Together, these contributions provide new infrastructure for pedagogy, computational musicology, and future MIR research on microtonal oud music.","primarySubject":"Computational musicology -> computational ethnomusicology","secondarySubject":"Analysis tasks -> indexing and querying;Applications -> digital libraries and archives;Applications -> music training and education;Cognitive and user-centered MIR -> music interfaces and services","presentation":"Virtual","slackChannel":"p6-8-oudtabs-an-interactive","channelUrl":"https://ismir2026.slack.com/archives/C0C60F26CG1","funFact":"The author, despite being a computer scientist, is known more as an oud player.","flags":["award"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_198.html","displayId":"P6-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Computational musicology","Computational ethnomusicology","Analysis tasks","Indexing and querying","Applications","Digital libraries and archives","Music training and education","Cognitive and user-centered MIR","Music interfaces and services","Awards Nominee","Open Review"]},"199":{"uid":"199","title":"Full-page Recognition of Handwritten Jeongganbo music scores: Dataset and Approach","authors":"Carlos Penarrubia; Danbinaerin Han; Dasaem Jeong; Jose J Valero-Mas","abstract":"Recognized as the first East Asian system to jointly encode pitch and rhythm, Jeongganbo notation represents a landmark in musicology and has been used in the Korean musical tradition since the 15th century. As a result, a vast amount of music content exists only in physical sheet form and is currently threatened by deterioration. Digitisation efforts have successfully archived these scores as digital images, but the lack of machine-readable transcriptions restricts their use by computational tools. Although Optical Music Recognition (OMR) offers a solution to automate this transcription, existing research on Jeongganbo has relied exclusively on software-rendered scores, which fail to capture the complexities of real documents. To address this limitation, this work presents two main contributions: (i) a new dataset comprising five handwritten Jeongganbo score books and 403 pages, along with their corresponding annotations for OMR tasks; and (ii) a two-step Deep Learning pipeline for full-page recognition of these scores. The results of a rigorous cross-domain evaluation of the methodology studied show competitive performance, while also providing useful insights into the design choices of Jeongganbo OMR systems. This approach marks a significant step towards bridging the gap between historical physical archives and modern digital accessibility.","primarySubject":"Analysis tasks -> optical music recognition","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"Virtual","slackChannel":"p6-9-full-page-recognition","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5W69U7","funFact":"","flags":["student"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_199.html","displayId":"P6-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Optical music recognition","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Open Review"]},"201":{"uid":"201","title":"Physiology-Driven Music Retrieval Using Contrastive Learning","authors":"Ananth Venkatesh; Elaine Chew","abstract":"Physiological signals provide a direct view of how music affects the body, yet remain underexplored as a retrieval modality. Prior work largely maps such signals onto low-dimensional valence–arousal representations, which are subjective and limited as ground truth. We frame physiology-driven music retrieval as a cross-modal contrastive learning task, investigating whether stimulus identity can be recovered directly from electrocardiographic traces and respiration signals of 110 participants listening to 30 classical piano performances. Both physiological and acoustic modalities are summarised using distributional statistics with participant-wise normalisation. Analysis using Earth Mover’s Distance shows true stimulus–response pairs align more strongly than random pairings (88.7% vs 50%, p < 0.001), indicating that the response contains a track specific signature within participants. A contrastive model trained with InfoNCE generalises this across participants to achieve Recall@1 = 0.662, Recall@2 = 0.909, and MRR = 0.814 in a 9-way retrieval task (p < 0.005). Reverse (music-to-physiology) retrieval collapses to chance, indicating shared stimulus-specific responses rather than individual responses. Despite the music features inducing a modest influence, explaining 14.1% of the variance in the physiological response, the effect is consistent across individuals and is sufficient to support cross-modal retrieval directly using physiological queries.","primarySubject":"Cognitive and user-centered MIR -> user behavioral and physiological responses","secondarySubject":"Applications -> music and health, well-being and therapy;Applications -> music retrieval systems;Cognitive and user-centered MIR -> perception and cognition;Cognitive and user-centered MIR -> personalization;Musical features and properties -> representation learning","presentation":"In-person","slackChannel":"p6-10-physiology-driven-music","channelUrl":"https://ismir2026.slack.com/archives/C0C60F0LRKK","funFact":"Ananth Venkatesh works on understanding music, with respect to the effect it has on the cardiovascular system. He enjoys all genres of music ranging from Indian Classical music to  Heavy Metal and is always open to sharing and receiving new music recommendations!","flags":["student"],"session":"PS-06","posterUrl":"https://ismir2026program.ismir.net/poster_201.html","displayId":"P6-10","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","User behavioral and physiological responses","Applications","Music and health, well-being and therapy","Music retrieval systems","Perception and cognition","Personalization","Musical features and properties","Representation learning","Open Review"]},"202":{"uid":"202","title":"Equivariant Music Transformer","authors":"Zixun Guo; Simon Dixon","abstract":"Humans recognize a musical passage even when it is shifted in time or transposed in pitch, indicating a notion of equivariance in the representation space. Our analysis, however, shows that standard music transformers map such time-shifted or pitch-transposed inputs onto uncorrelated representations: these models become progressively less equivariant as they scale in size or train longer. This suggests that in standard music transformers, additional model capacity is allocated to memorizing absolute patterns rather than capturing shared musical structures. In this paper, we propose the Equivariant Music Transformer (EMT), which enforces equivariance through self-distillation by jointly optimizing a next-token-prediction and an auxiliary equivariance regularization loss. We find that the additional equivariance loss acts as a beneficial regularizer, simultaneously improving next-token prediction and producing equivariant latent representations. Through both objective and subjective evaluations, EMT demonstrates superior equivariance and generative capability compared to data augmentation, feature engineering, and state-of-the-art (SOTA) baselines. More broadly, our findings reveal that standard language modeling methods alone do not capture music’s translational symmetries, and dedicated inductive biases are required to produce better music representations. The code, weights and demos are available online.","primarySubject":"Musical features and properties -> representations of music","secondarySubject":"MIR data fundamentals -> symbolic music processing","presentation":"In-person","slackChannel":"p7-1-equivariant-music-transformer","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7WRZQD","funFact":"Zixun Nicolas Guo is a guitar player.\nYou might recognize him from the jam session (ISMIR 25) where he played solo guitar. \nHe also presented GAPS during ISMIR 24.","flags":["student","award"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_202.html","displayId":"P7-01","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representations of music","MIR data fundamentals","Symbolic music processing","Awards Nominee","Open Review"]},"237":{"uid":"237","title":"Long, Short, and In Between: Non-Isochronous Models of Timing in Balkan Song Performances","authors":"George Sioros","abstract":"Musical timing is often shaped by systematic non-isochronous duration patterns. Computational and music-theoretical models recognize such timing as fundamental to many musical styles and typically represent it either as deviations from an isochronous reference structure or through statistical description. These approaches are effective for describing observed timing, but they provide limited explicit structure and parametrization for modeling local variation. Non-Isochronous (NI) Grids offer an alternative by modeling timing as patterns of two beat durations, Long and Short, that capture local organization and cycle-to-cycle change through a compact set of parameters. In this paper, we test whether NI Grids can account for timing observed in Aksak rhythms from Balkan songs. Using a fitting process constrained by the music-theoretical organization of Aksak rhythms, we model previously reported inter-beat interval sequences from two songs with RMS residual errors approximately 10 ms. Most cycles in each song are modelled by one Long/Short configuration, with variation captured through changes in the duration ratio. A few cycles require different but closely related configurations for comparable residual errors. Because several such configurations can fit each cycle, distinguishing among them would require additional ethnomusicological and theoretical constraints, and onset information from between-beat events.","primarySubject":"Computational musicology -> computational music theory and musicology","secondarySubject":"Computational musicology -> computational ethnomusicology;Computational musicology -> systematic musicology;Musical features and properties -> expression and performative aspects of music;Musical features and properties -> representations of music;Musical features and properties -> rhythm, beat, tempo","presentation":"Virtual","slackChannel":"p7-2-long-short-and","channelUrl":"https://ismir2026.slack.com/archives/C0C60F3HEDP","funFact":"","flags":[],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_237.html","displayId":"P7-02","presenterMode":"online","materials":["paper"],"keywords":["Computational musicology","Computational music theory and musicology","Computational ethnomusicology","Systematic musicology","Musical features and properties","Expression and performative aspects of music","Representations of music","Rhythm, beat, tempo"]},"250":{"uid":"250","title":"Unified Music Identification for Tracks and Versions","authors":"R. Oguz Araz; Joan Serrà; Yuki Mitsufuji; Xavier Serra; Dmitry Bogdanov","abstract":"Given a music database, track identification (TI) retrieves the exact track matching an audio excerpt, whereas version identification (VI) retrieves its musical versions. Traditionally, the two tasks have been addressed separately. However, as every track is its own closest version, we investigate whether VI can subsume TI. This requires VI systems to be robust to both signal manipulation and audio degradation. We therefore propose a unified benchmark that evaluates accuracy and robustness on each task. Comparing seven existing models on this benchmark, we show that none of them are both accurate and robust on both tasks. We then train a baseline model targeting both tasks and show that a unified system is possible with 10 s TI queries. Lastly, we characterize the two retrieval constraints that limit our model’s TI performance. We envision extending this unification to other music identification tasks.","primarySubject":"Applications -> music retrieval systems","secondarySubject":"Analysis tasks -> alignment, synchronization, and score following;Analysis tasks -> indexing and querying;Analysis tasks -> pattern matching and detection;Analysis tasks -> similarity metrics;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p7-3-unified-music-identification","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7XR8JZ","funFact":"","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_250.html","displayId":"P7-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications","Music retrieval systems","Analysis tasks","Alignment, synchronization, and score following","Indexing and querying","Pattern matching and detection","Similarity metrics","Musical features and properties","Representations of music","Open Review"]},"256":{"uid":"256","title":"When AI Annotates First: Measuring Its Impact on Music Score Labeling","authors":"Alejandro Galan-Cuenca; Juan P. Martinez-Esteso; Carlos Pérez-Sancho; Francisco J. Castellanos; Antonio Javier Gallego","abstract":"Pixel-level annotation of music score images is essential for training reliable Optical Music Recognition and document analysis systems, but it remains a time-consuming task that requires expert intervention. This paper presents PixelSmart, an open-source AI-assisted annotation tool for music score layout analysis. Instead of annotating from scratch, users start from model-generated proposals and focus on correcting and validating them. The tool also integrates reliability indicators and supports human-in-the-loop fine-tuning from corrected annotations. We evaluate the proposed workflow with five annotators on three music score datasets, comparing it against a general-purpose editor and a specialized manual annotation tool. Results show that AI-assisted annotation improves both efficiency and final label quality. PixelSmart reaches 90.2% F1 with out-of-domain initialization and 92.0% with in-domain initialization, computed against the original ground truth, outperforming the manual baselines, which obtain 88.2% and 87.9%. In the best setting, annotation time is reduced to 3:02 min. per patch, corresponding to reductions of 71% and 63% with respect to the manual baselines. Moreover, after only five corrected samples, fine-tuning reduces the out-of-domain annotation time from 7:09 to 3:41 min. per patch. These results show that AI-first annotation can make music score corpus creation more accurate, efficient, and scalable.","primarySubject":"Analysis tasks -> optical music recognition","secondarySubject":"Analysis tasks;Applications","presentation":"Virtual","slackChannel":"p7-4-when-ai-annotates","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5XLQT1","funFact":"No animals were harmed in the making of this work \nAll researchers were rewarded with treats\nOur annotation tool is available here: https://github.com/cperezs/pixelsmart","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_256.html","displayId":"P7-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Optical music recognition","Applications","Open Review"]},"266":{"uid":"266","title":"Revisiting Two-Stage Audio Chord Recognition in the Deep-Learning Era: Leveraging Multi-Pitch Estimation and Symbolic Pretraining","authors":"Yiwei Ding; Yannik Venohr; Sebastian Strahl; Meinard Müller; Christof Weiß","abstract":"In the deep-learning era, automatic chord recognition (ACR) from audio predominantly relies on end-to-end strategies. By contrast, most early work on audio ACR followed a two-stage paradigm: first extracting a pitch-based representation from audio, and then predicting chord labels from the representation. In this paper, we revisit the two-stage paradigm in the deep-learning era by explicitly decomposing audio ACR into multi-pitch estimation (MPE) from audio followed by chord recognition based on the multi-pitch representation. We leverage existing large-scale transcription datasets to pretrain the MPE module and chord-annotated symbolic music datasets to pretrain the chord recognition module, thereby addressing the scarcity of annotated audio data. This allows the MPE module to handle diverse instrumentation and acoustic conditions, and the chord recognition module to capture harmonic diversity. Experiments show that two-stage models consistently outperform end-to-end baselines on Western classical music and that fine-tuning on popular music outperforms end-to-end models trained directly on pop music, demonstrating cross-genre transferability. Our work establishes the two-stage approach as a robust and effective framework for deep-learning-based audio ACR.","primarySubject":"Musical features and properties -> harmony, chords and tonality","secondarySubject":"Analysis tasks;MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p7-5-revisiting-two-stage","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7YG5NH","funFact":"Automatic Chord Recognition is one of the most difficult research topics, the author claims, because it is so well-studied.\n\nThe paper is based on part of a submission rejected by ICASSP. The other parts... Let's see where I will make them!","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_266.html","displayId":"P7-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Harmony, chords and tonality","Analysis tasks","MIR data fundamentals","Music signal processing","Open Review"]},"195":{"uid":"195","title":"The Internet Archive Music Dataset","authors":"Paraskevas Stamatiadis; Bernardo V. Miranda; Clémentine Berger; Gaël Richard; Mathieu Fontaine; Slim Essid","abstract":"We introduce the Internet Archive Music Dataset (IAMD), a large-scale collection of captioned music segments derived from the Internet Archive. To the best of our knowledge, IAMD constitutes the largest publicly available music–caption dataset to date with over 34,000 hours of audio, providing a valuable benchmark for training and evaluating music understanding and generative models. The dataset is built from content declared to be distributed under Creative Commons licenses, and cross-referencing with MusicBrainz is done to improve license information reliability. To annotate IAMD, we present an automatic captioning pipeline that augments base captions produced by an audio-language model (ALM) with textual metadata sourced from the Internet Archive and imputed metadata obtained using audio classification models. Caption quality is assessed objectively and subjectively, and results indicate that the annotation pipeline is reliable and does not degrade caption quality with scaling.","primarySubject":"Evaluation, datasets, and reproducibility","secondarySubject":"Evaluation, datasets, and reproducibility -> annotation protocols;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p7-6-the-internet-archive","channelUrl":"https://ismir2026.slack.com/archives/C0C60F1SMED","funFact":"All authors of the paper play at least 1 musical instrument.","flags":[],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_195.html","displayId":"P7-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Annotation protocols","Novel datasets and use cases","Generative Tasks","Music and audio synthesis"]},"280":{"uid":"280","title":"Robust Instrument-Agnostic Music Transcription for Western Classical Music","authors":"Yannik Venohr; Christof Weiß","abstract":"Musicological corpus analysis from audio, among other scenarios, requires Automatic Music Transcription (AMT) methods that generalize across instrumentations and acoustic conditions. While AMT has advanced substantially, existing methods lack robustness: a model may transcribe high-quality recordings well but fail on noisy historical ones, or have problems with challenging instrumentations such as choir. To study such behavior, we focus on Western classical music as an application domain and compile a large, heterogeneous data scenario from existing datasets. We perform cross-dataset evaluation on seven held-out test sets, including a recently introduced robustness benchmark. As our main contribution, we introduce ClassicPitch (CP), a CNN model for instrument-agnostic music transcription on the frame and note levels. Rather than introducing a single novel component, CP combines and modernizes established techniques into a coherent and scalable framework: a convolutional frame-level backbone with efficient ConvNeXt blocks, and onsets derived from a shared frame representation. In a variant with under 50K parameters, our model is already superior to substantially larger models on various datasets and evaluation metrics, demonstrating that competitive and robust instrument-agnostic AMT is achievable with a lightweight fully convolutional model. More importantly, CP remains stable on noisy real-world recordings. We publish CP as an open-source, open-weight tool for musical corpus analysis and beyond.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology","presentation":"Virtual","slackChannel":"p7-7-robust-instrument-agnostic","channelUrl":"https://ismir2026.slack.com/archives/C0C627DFMKN","funFact":"The author has is practicing to transcribe music himself. So far, ClassicPitch is learning faster than he is.\n\nClassicPitch has already found work outside this paper: check out Yiwei Ding's paper on chord recognition at this ISMIR to see it in action.\n\nThe author wasn't entirely sure how much fun is legally allowed in a fun fact. To be safe, he left plenty of headroom.","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_280.html","displayId":"P7-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Evaluation, datasets, and reproducibility","Evaluation methodology","Open Review"]},"534":{"uid":"534","title":"\"I Just Kept Clicking Until It Hit\": Exploring Virtual Instrument Search behavior and Challenges Across User Groups","authors":"Subeen Kim; Eunsik Shin; Woojae Cho; Kyogu Lee","abstract":"Virtual instrument search is a common yet relatively underexplored bottleneck in music production. Although users often have a clear idea of the sound they want, they differ substantially in how they describe target sounds, navigate preset libraries, and evaluate search results. To better understand these differences, we conducted a mixed-methods study with 33 participants who completed preset-retrieval tasks in Logic Pro's Alchemy followed by semi-structured interviews. Our analysis shows that users can be meaningfully characterized along two dimensions—music domain knowledge and search experience—and that they form four distinct groups for understanding instrument-search users. Across groups, we observed systematic differences in verbal expression, search behavior, perceived system limitations, and desired technical, functional, and interface support. Furthermore, we explored the potential of utilizing free-form text inputs to capture how users' personalized backgrounds uniquely shape their descriptions of virtual instruments. These results suggest that instrument-retrieval systems should support diverse forms of expression and search pathways.","primarySubject":"Cognitive and user-centered MIR","secondarySubject":"Applications -> music retrieval systems;Cognitive and user-centered MIR -> human-computer interaction;Cognitive and user-centered MIR -> music interfaces and services;Cognitive and user-centered MIR -> perception and cognition;Cognitive and user-centered MIR -> personalization","presentation":"","slackChannel":"p7-8-i-just-kept","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6PHZ2T","funFact":"1. I’m a hamster lover.\n2. I have a quirky habit when I get sleepy\n3. I only get into bed after a shower. If I’m too sleepy to shower, I sleep on the floor instead.","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_534.html","displayId":"P7-08","presenterMode":"","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","Applications","Music retrieval systems","Human-computer interaction","Music interfaces and services","Perception and cognition","Personalization","Open Review"]},"293":{"uid":"293","title":"MEX: The Musical Expectation Benchmark Suite","authors":"Mathias Rose Bjare; Augustin Bouquillard; Giorgia Cantisani; Seung-Goo Kim; Stefan Lattner; Gerhard Widmer","abstract":"We propose the Musical Expectation (MEX) benchmark suite for perceptually validating computational models of musical expectation, i.e., models that estimate music information dynamics (MID) measures of perceived musical expectations, such as surprisal or uncertainty from symbolic music or musical audio.\nMEX collects and extends tasks for evaluating such models previously proposed in music information retrieval and cognitive sciences using standard datasets, and evaluates whether:\n(1) proposed MID measures decrease with repetition due to in-context learning of the computational model,\n(2) peaks in such measures align with boundaries of music segments,\n(3) pitch surprisal and uncertainty estimated from audio are consistent with those modeled in event-based symbolic music,\n(4) proposed MID measures can predict electrophysiological responses to music.\nWe stress that this benchmark suite does not provide an exhaustive computational characterization of music expectation.\nRather, we envision this as an active, interdisciplinary initiative, with benchmarks evolving as our understanding of the underlying perceptual processes deepens. \nWe provide open-access code to estimate surprisal and uncertainty from audio using baseline models, run the tasks on standard datasets, and include utilities for running the tasks with users' own data.","primarySubject":"Cognitive and user-centered MIR -> perception and cognition","secondarySubject":"Cognitive and user-centered MIR -> user behavioral and physiological responses;Evaluation, datasets, and reproducibility;Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Evaluation, datasets, and reproducibility -> reproducibility","presentation":"Virtual","slackChannel":"p7-9-mex-the-musical","channelUrl":"https://ismir2026.slack.com/archives/C0C6STMPYTA","funFact":"","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_293.html","displayId":"P7-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","Perception and cognition","User behavioral and physiological responses","Evaluation, datasets, and reproducibility","Evaluation methodology","Novel datasets and use cases","Reproducibility"]},"523":{"uid":"523","title":"Understanding Automatic Mixing: A Subtask-Oriented Analysis of Two-Stage Mixing System","authors":"Jinjie Shi; Wei Hua; Kunzhu Xie; Make Li; Yuchen Liu; Joshua Reiss","abstract":"Automatic mixing transforms multitrack recordings into perceptually coherent, balanced, and aesthetically consistent mixes. However, in real-world production, this task is challenging due to large track counts, diverse instrumentation, and strong inter-track dependencies. Two-stage systems address this complexity by separating intra-group processing from inter-group mixing, yet it remains unclear whether their gains arise from stronger component models or from explicit task decomposition. We present a subtask-oriented analysis of automatic mixing through three controlled listening experiments. We investigate whether full-mix models transfer to intra-group mixing, whether downstream models compensate for grouping and loudness errors, and whether two-stage decomposition improves full-mix quality. Across three dense pop and rock excerpts, transfer differs between the evaluated models; inappropriate grouping causes clear downstream degradation, while altered loudness relationships have weaker and model-dependent effects. Both two-stage variants significantly outperform their corresponding single-stage baselines. These findings support explicit separation of local balance and global mix coordination as a useful design principle for automatic mixing. Code and audio examples are available online.","primarySubject":"Applications -> music composition, performance, and production","secondarySubject":"Generative Tasks -> qualitative evaluations;MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p7-10-understanding-automatic-mixing","channelUrl":"https://ismir2026.slack.com/archives/C0C5YD4JPDG","funFact":"Automatic mixing has a 50-year research history. Dan Dugan published “Automatic Microphone Mixing” in 1975, one of the field’s earliest foundational papers.\n\nIs loudness normalization always fair? If a system can produce a louder audio without introducing artifacts, that capability may itself be part of its mixing performance.\n\nCan a training-free rule-based system match state-of-the-art deep learning models on a subtask?","flags":["student"],"session":"PS-07","posterUrl":"https://ismir2026program.ismir.net/poster_523.html","displayId":"P7-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications","Music composition, performance, and production","Generative Tasks","Qualitative evaluations","MIR data fundamentals","Music signal processing","Open Review"]},"448":{"uid":"448","title":"Adapting Music Foundation Models to Iranian Classical Music","authors":"Pouya Mohseni; Bagher BabaAli; Ichiro Fujinaga","abstract":"This work evaluates whether Western-trained music foundation models transfer effectively to the specialized domain of Iranian Classical Music, and compares four adaptation regimes: probing, full fine-tuning, parameter-efficient transfer learning, and continued pre-training. Using a 620-hour corpus of domain-specific audio, we systematically study how much adaptation is required for transfer and when domain-specific pre-training yields meaningful gains. These approaches are benchmarked across five downstream tasks. Our results show that while parameter- and data-efficient methods achieve strong performance on supervised tasks, they exhibit reduced transferability to unseen tasks within the same domain compared to more computationally intensive strategies. Furthermore, we find that continued pre-training is not strictly necessary for competitive downstream performance, although it can improve cross-task transferability without introducing additional inference overhead.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks -> automatic classification;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Musical features and properties -> representations of music;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"In-person","slackChannel":"p8-1-adapting-music-foundation","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCWDRH8","funFact":"","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_448.html","displayId":"P8-01","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Automatic classification","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Representations of music","Timbre, instrumentation, and singing voice","Open Review"]},"458":{"uid":"458","title":"SmartLooper: a Framework for Personalized Human-AI Musical Improvisation","authors":"Finlay Miller; Sageev Oore; Chandramouli Sastry; Marvin da Silva; Sri Harsha Dumpala; Dani Oore; Scott Lowe","abstract":"Real-time human--AI musical collaboration presents a persistent design challenge: AI contributions tend to be either too conservative to be interesting or too unpredictable to play along with. We posit that looping (along with related practices of vamping and grooving) offers a framework that sidesteps this trade-off. These recurring structures bound variation within forms the musician already understands, and let development accrue over musical rather than sample-level timescales---a balance upon which improvisers have long relied when playing with each other.\n\nWe present SmartLooper, a system for real-time human–AI piano improvisation built on this premise. The performer first assembles a personal corpus of their own loop-based recordings, giving them complete control over the stylistic space the system inhabits. SmartLooper segments this corpus into short loops and embeds them using pretrained MIDI models. In performance, the system improvises by retrieving loops from the corpus via a stochastic walk in embedding space: successive loops are chosen for proximity so that transitions feel continuous.\n\nTo broaden the corpus without diluting its character, we augment the data with synthetic loops in the style of the corpus generated using MIDI-GPT. These are filtered down to only high-quality samples using a discriminator trained to distinguish degraded MIDI from samples in the personal corpus. The resulting dataset extension provides broader support for fluid transitions while preserving the performer's voice.\n\nSmartLooper has been used by professional musicians in over ten hours of live performance for audiences totalling roughly a thousand listeners so far; we provide example recordings.","primarySubject":"Creativity -> creative practice involving MIR or generative technology","secondarySubject":"Analysis tasks -> similarity metrics;Applications -> music composition, performance, and production;MIR data fundamentals -> symbolic music processing","presentation":"Virtual","slackChannel":"p8-2-smartlooper-a-framework","channelUrl":"https://ismir2026.slack.com/archives/C0C627XH7HS","funFact":"Finlay doesn't actually know how to play the piano.","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_458.html","displayId":"P8-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Creativity","Creative practice involving MIR or generative technology","Analysis tasks","Similarity metrics","Applications","Music composition, performance, and production","MIR data fundamentals","Symbolic music processing"]},"468":{"uid":"468","title":"D-Composer: Language Modeling for Simultaneous Drum Transcription and Sound Event Decomposition","authors":"Patrick O'Reilly; Saumya Pailwan; Annie Chu; Jason Smith; Bryan Pardo","abstract":"Drum beats are often conceptualized and produced as arrangements of repeated single-drum sound events or one-shots. Drum Analysis-by-Synthesis (AbS) models combine this structural assumption with reconstruction-based training to perform MIR tasks such as drum transcription, drum source separation, and one-shot extraction, offering a flexible alternative to more traditional task-specific supervision. However, leading AbS approaches struggle to generalize to a broad range of drum timbres (e.g., electronic drum kits), rely on high-quality paired audio-transcription training data, and do not model intra-class timbral variation within a single arrangement (e.g., multiple kick sounds). To overcome these limitations, we propose D-Composer, an autoregressive language modeling framework that decomposes drum recordings into labeled sound events and their onset times. In drum transcription and one-shot extraction tasks, D-Composer performs comparably to or better than a state-of-the-art AbS system and task-specific baselines while generalizing to a broader range of drum timbres. Our results demonstrate that when equipped with abundant synthetic data and compact acoustic token representations, language modeling provides a simple recipe for unifying transcription and sound event decomposition.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Analysis tasks -> sound source separation;Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p8-3-d-composer-language","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCWFWHG","funFact":"Treats is the debut studio album by the American noise pop duo Sleigh Bells, consisting of the vocalist Alexis Krauss and the producer-guitarist Derek Miller.\nTreats was released on May 24, 2010, by Mom+Pop and N.E.E.T. Recordings.\nThe release of Treats was preceded by the April single \"Tell 'Em\" and featured the group's most popular track, the Funkadelic-sampling \"Rill Rill\".","flags":["student","award"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_468.html","displayId":"P8-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Sound source separation","Generative Tasks","Music and audio synthesis","Awards Nominee","Open Review"]},"315":{"uid":"315","title":"CoALa: A Plugin Model for Controlling Audio Latents","authors":"Sarah Nabi; Nabarun Goswami; Philippe Esling; Geoffroy Peeters; Frédéric Bevilacqua; Tatsuya Harada","abstract":"In recent years, neural audio synthesis has advanced significantly, enabling high-quality real-time waveform generation on standard CPUs. However, defining personalized controls for live musical performance remains challenging, as existing methods typically require large labeled datasets and rely on heavy architectures that are generally impractical for real-time use. In this paper, we propose CoALa, a lightweight plugin model for Controlling Audio Latents in real-time, which enables users to add time-varying musical controls on pretrained neural audio synthesis models using only a limited set of labeled examples. Our method leverages a variational autoencoder with a contrastive learning regularization to disentangle control-relevant features from the pretrained latent space in designated dimensions. We rely on labeled frames as an augmentation strategy to generate multiple views of similar control variations for each unlabeled latent sequence, and apply the contrastive objective on these variation embeddings. By augmenting the pretrained latent space with temporal latent tokens, our method supports an arbitrary number of control variables, generalizes to highly compact latent representations, and can simultaneously disentangle multiple time-varying controls evolving at different temporal scales. Our model defines a bidirectional mapping between latent and control trajectories and achieves high control accuracy while preserving synthesis quality for feature extraction, editing, and conditional synthesis tasks, using as little as 10% labeled data. Its lightweight streamable architecture further enables fast CPU inference, making it well suited for real-time applications in Max/MSP environments. Our code is available at https://github.com/nabsara/coala.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Creativity -> tools for artists;Generative Tasks -> real-time considerations;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p8-4-coala-a-plugin","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5YU7ST","funFact":"You might recognize us from last year ISMIR 2025 where we presented \"PLaTune: Adding temporal musical controls on top of pretrained generative models\".","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_315.html","displayId":"P8-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Creativity","Tools for artists","Real-time considerations","Musical features and properties","Representation learning","Open Review"]},"317":{"uid":"317","title":"MAJEPPA: Morphing and Assessing in a Unified Piano Performance Space","authors":"Jinwen Zhou; Huan Zhang; Weixi Zhai; Jinhua Liang; Aidan O. T. Hogg; Simon Dixon","abstract":"We present MAJEPPA, a self-supervised framework to learn piano performance representations that span the full skill spectrum, from beginner practice sessions to virtuoso concert recordings. We curate the MAJEPPA dataset, comprising ~4,000 annotated recordings across six expertise levels and six recording contexts. We adapt a single pre-trained MIDI autoregressive model with a joint objective: next-token prediction learns score-conditioned performance generation at various skill levels, while InfoNCE and supervised contrastive losses align abstract score and performance representations in a joint embedding space. The proposed model both generates and understands performances in a unified framework. By introducing the EVPMR benchmark, a suite of downstream tasks spanning quality assessment, competition ranking, mistake and technique classification, we evaluate the learnt representations, demonstrating progress towards a real-world model for the piano performance space.","primarySubject":"Musical features and properties -> expression and performative aspects of music","secondarySubject":"Applications -> music training and education;Evaluation, datasets, and reproducibility -> evaluation metrics;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Generative Tasks -> artistically-inspired generative tasks;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p8-5-majeppa-morphing-and","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7ZM23X","funFact":"","flags":["student","award"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_317.html","displayId":"P8-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Expression and performative aspects of music","Applications","Music training and education","Evaluation, datasets, and reproducibility","Evaluation metrics","Novel datasets and use cases","Generative Tasks","Artistically-inspired generative tasks","Representation learning","Awards Nominee","Open Review"]},"384":{"uid":"384","title":"Harmonic Representations in Symbolic Music Transformers are Localized and Controllable","authors":"Adi Silberschein; Megan Wei; Tamar Rott Shaham","abstract":"Transformer models trained on symbolic music have become remarkably capable. However, it is yet unclear how they represent music structure, and whether understanding their internal representations can be used to unlock methods from model control. We examine this through a case study of harmonic analysis: when a music transformer navigates a cadence, what does it internally ``know'' about its harmonic context, and can that knowledge be used to affect the generation? Using Aria, a LLaMA-based symbolic music transformer evaluated on Bach chorales, we find that harmonic decisions are causally localized progressively from middle layers onward. Linear probes show that mode becomes decodable in the same layer range where patching identifies growing causal effect, while key identity emerges later. We then confirm these representations are actionable for music generation control: steering via latent directions corresponding to harmonic changes reliably produces a deceptive resolution, while preserving structural coherence and generation quality. This demonstrates that, at inference time, the model's internal harmonic knowledge can be directly manipulated to control the harmonic output.","primarySubject":"Musical features and properties","secondarySubject":"Computational musicology -> computational music theory and musicology;MIR data fundamentals -> symbolic music processing;Musical features and properties -> harmony, chords and tonality;Musical features and properties -> representation learning;Musical features and properties -> representations of music","presentation":"","slackChannel":"p8-6-harmonic-representations-in","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJH3ZSS","funFact":"Adi, the paper's first author, is also a singer-songwriter. Her single Oh My Frankenstein is sung from the point of view of Frankenstein's creature, pleading with the scientist who built it and then walked away. Read \"creature\" as \"AI\" if you like. Video: https://www.youtube.com/watch?v=Aox1twWxvXo · Spotify: https://open.spotify.com/album/5Ou24pFvnaZdalXjJ5Q92H","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_384.html","displayId":"P8-06","presenterMode":"","materials":["paper","poster","video"],"keywords":["Musical features and properties","Computational musicology","Computational music theory and musicology","MIR data fundamentals","Symbolic music processing","Harmony, chords and tonality","Representation learning","Representations of music","Open Review"]},"342":{"uid":"342","title":"From Prediction to Collaboration: Interactive Symbolic Music Analysis","authors":"Emmanouil Karystinaios; Johannes Hentschel; Markus Neuwirth; Gerhard Widmer","abstract":"Automatic symbolic music analysis has made substantial progress, yet existing systems are typically designed for a single mode of use, such as full-score prediction, and therefore do not match the broader range of operations that arise in analysis workflows, including partial completion, local correction, and iterative refinement. As a result, there remains a gap between strong benchmark models and systems that can support interactive analytical use. We present a unified framework for symbolic Roman-numeral (RN) analysis that narrows this gap by combining strong predictive performance with direct support for constrained completion and revision. The method is designed to provide a practical trade-off between accuracy and interactive responsiveness by computing expensive pretrained representations once and reusing them during iterative refinement, making powerful pretrained models more amenable to interactive settings. It supports complete score analysis, targeted revision of existing labels, and inference of missing annotations from partial context through a shared modeling framework. Experiments on Dilemmadata, the largest and most heterogeneous benchmark of its kind, show that the proposed approach is a strong RN-analysis baseline while also supporting masked completion from partial labels. Together with a prototype interface for multi-level candidate inspection and editing, these results position automatic RN analysis not only as a prediction problem, but also as a foundation for future interactive tools for music analysis.","primarySubject":"Musical features and properties -> harmony, chords and tonality","secondarySubject":"Analysis tasks;Cognitive and user-centered MIR -> human-computer interaction;Computational musicology -> computational music theory and musicology;MIR data fundamentals -> symbolic music processing","presentation":"Virtual","slackChannel":"p8-7-from-prediction-to","channelUrl":"https://ismir2026.slack.com/archives/C0C60F6GA1F","funFact":"1) Since this paper was published, first author Emmanouil Karystinaios has joined MuseScore, while second author Johannes Hentschel has moved to TU Dortmund as an Associate Professor.\n2) AnalysisGNN is already making a second ISMIR 2026 appearance, powering the Late-Breaking Demo “ScorePrompts.”\n3) This is not just a paper demo: you can try the model and correction workflow yourself at analysisgnn.com","flags":[],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_342.html","displayId":"P8-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Harmony, chords and tonality","Analysis tasks","Cognitive and user-centered MIR","Human-computer interaction","Computational musicology","Computational music theory and musicology","MIR data fundamentals","Symbolic music processing"]},"351":{"uid":"351","title":"A First Benchmark and Metric for Evaluating Polyphony in Symbolic Music using Voice Independence","authors":"Eleanor Row; György Fazekas","abstract":"Polyphony in symbolic music is often measured by vertical note density or other basic objective metrics. Such metrics, however, fail to account for voice independence and counterpoint textures that are central to musicology. We introduce a compact, musically informed benchmark dataset for evaluating multi-part polyphony: Polyphony-Ranked Excerpt Dataset (PolyRED), alongside a proposed metric: the Polyphonic Voice Independence Index (PVII). PolyRED comprises rankings of 15 four-part excerpts from three domains: Bach chorales, classical and modern string quartets, annotated by 25 score-literate musicians. Agreement was moderate across the evaluation trials (mean Kendall’s W=0.58, rising to 0.64 in sensitivity analysis). These rankings were then used to derive pairwise win-rate reference scores, serving as human-derived benchmark targets for evaluating new metrics. PVII quantifies independence through similarity between melodic, rhythmic, and onset-based part representations. On this compact benchmark, PVII produced rankings that were more closely aligned with aggregated score-literate participant rankings than density- and overlap-based baselines. These results provide initial evidence that voice-independence features capture aspects of perceived polyphony in short, four-part, predominantly monophonic symbolic textures.","primarySubject":"Evaluation, datasets, and reproducibility -> evaluation metrics","secondarySubject":"Computational musicology -> computational music theory and musicology;Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> novel datasets and use cases;MIR data fundamentals -> symbolic music processing;Musical features and properties -> melody and motives","presentation":"In-person","slackChannel":"p8-8-a-first-benchmark","channelUrl":"https://ismir2026.slack.com/archives/C0C627JJUR2","funFact":"Eleanor Row plays the violin in a Tango Orchestra in London! \nYou might recognise George from his many years attending ISMIR! Next year he will help to host it in London.\nIf you like talking about musical textures, come catch Eleanor at the poster and pick her brains on heterophony!","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_351.html","displayId":"P8-08","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Evaluation metrics","Computational musicology","Computational music theory and musicology","Evaluation methodology","Novel datasets and use cases","MIR data fundamentals","Symbolic music processing","Musical features and properties","Melody and motives"]},"303":{"uid":"303","title":"CapCyCin: Improving Music Captioning through Cycle-Consistent Music–Text Alignment","authors":"Saeyeon Hwang; Kyungsu Kim; Kyogu Lee","abstract":"Music captioning is inherently ambiguous as there is no single correct description for a given music. This makes it difficult to define and evaluate what should be considered a good caption. Existing audio captioning metrics often fail to capture semantic equivalence across diverse captions and may not always provide accurate scores for semantically inconsistent descriptions. In addition, human evaluation is costly and subjective. To address these limitations, we propose CapCyCin, a Cycle-Consistency-based framework that improves music Captioning through music–text alignment. For a given music, we generate a caption and reconstruct the music via a text-to-music model. We then measure the similarity between the original and reconstructed audio. Using this cycle consistency score, our approach: (1) constructs a large-scale preference dataset of over 410K pairs by ranking candidate outputs, (2) trains a reward model for evaluating captioning quality and modeling human preference, and (3) enables Direct Preference Optimization (DPO) to improve captioning models without human annotation. Our approach improves music–text alignment, leading to more fine-grained music captions, and suggests that these improvements can further benefit both text-conditioned music generation and music captioning. We publicly release our dataset, model, and code at https://github.com/CapCyCin/CapCyCin/.","primarySubject":"MIR data fundamentals -> multimodality","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p8-9-capcycin-improving-music","channelUrl":"https://ismir2026.slack.com/archives/C0C60F4HGUR","funFact":"","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_303.html","displayId":"P8-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Multimodality","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Musical features and properties","Representation learning","Open Review"]},"372":{"uid":"372","title":"Geometric Iterative Retrieval for Neural Audio Codec Resynthesis","authors":"Leo Schmidt-Traub; Frederic Berdoz; Luca A. Lanzendörfer; Roger Wattenhofer","abstract":"Neural audio codecs based on Residual Vector Quantization (RVQ) have become the dominant discrete representation for token-based general audio generation, yet resynthesizing high-quality audio from coarse codec tokens remains an open problem. Prior work has framed resynthesis as a choice between discrete token prediction and continuous regression. We argue that this dichotomy is incomplete and introduce geometric iterative retrieval, a paradigm that uses the RVQ layer hierarchy itself as a natural iterative decomposition in continuous codebook space. Rather than classifying over discrete vocabularies or regressing to a single target vector, our method performs contrastive retrieval in the codebook's geometric space. We evaluate our approach on codec restoration tasks across speech and music, demonstrating improvements over both token prediction and one-step regression baselines.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"","presentation":"Virtual","slackChannel":"p8-10-geometric-iterative-retrieval","channelUrl":"https://ismir2026.slack.com/archives/C0C5T884K37","funFact":"","flags":["student"],"session":"PS-08","posterUrl":"https://ismir2026program.ismir.net/poster_372.html","displayId":"P8-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning"]},"205":{"uid":"205","title":"Learning Jazz Pianist Style with Cross-Attention Conditioning","authors":"Drew Edwards; Akira Maezawa; Simon Dixon","abstract":"Jazz pianists develop distinctive traits that experienced listeners can often identify within seconds, yet the features underlying this recognition resist formal description. We study jazz pianist style through the lens of a pretrained symbolic music transformer, showing that its learned representations already encode pianist identity well enough for highly accurate classification across two benchmarks. We then augment the transformer with cross-attention over learned pianist identity embeddings, enabling it to generate music conditioned on a specific artist's style. Two evaluation protocols confirm that the generator captures meaningful stylistic structure: a sliding-window classifier consistently attributes conditioned continuations to the correct artist, far above unconditioned baselines; and a classifier trained entirely on synthetic generations---never seeing real performances---identifies real pianists with 87% chunk-level and 96% song-level accuracy. Finally, we repurpose the classifier to locate the most characteristic moments within a performance, surfacing the specific musical gestures that distinguish each pianist's voice.","primarySubject":"Generative Tasks","secondarySubject":"Generative Tasks -> artistically-inspired generative tasks;Generative Tasks -> evaluation metrics;MIR data fundamentals -> symbolic music processing;Musical features and properties -> expression and performative aspects of music;Musical features and properties -> musical style and genre","presentation":"","slackChannel":"p9-1-learning-jazz-pianist","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7WE5EH","funFact":"1. The prompt behind the paper's listening demo (\"Ain't Misbehavin'\") was played by its first author, Drew Edwards. The model then continues his own opening in the style of each of the twelve pianists, so on the demo page he gets to hear his own playing \"continued\" by Art Tatum, Erroll Garner and ten others.\n\n2. Dick Hyman, whose 1994 book of études \"In the Styles Of... The Great Jazz Pianists\" inspired the project, is himself one of the twelve pianists modelled. When the classifier is unsure who a generated passage sounds like, its favourite fallback is Hyman: fitting for a pianist who made a career of playing in everyone else's style, but possibly more of an artifact of data imbalance.","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_205.html","displayId":"P9-01","presenterMode":"","materials":["paper","poster","video"],"keywords":["Generative Tasks","Artistically-inspired generative tasks","Evaluation metrics","MIR data fundamentals","Symbolic music processing","Musical features and properties","Expression and performative aspects of music","Musical style and genre","Open Review"]},"13":{"uid":"13","title":"Streaming Generation for Music Accompaniment","authors":"Yusong Wu; Mason Wang; Heidi Lei; Stephen Brade; Lancelot Blanchard; Shih-Lun Wu; Aaron Courville; Anna Huang","abstract":"Music generation models can produce high-fidelity coherent accompaniment given complete audio input, but are limited to editing and loop-based workflows. We study real-time audio-to-audio accompaniment: as a model hears an input audio stream (e.g., vocals), it has to also simultaneously generate in real-time a coherent accompanying stream (e.g., guitar). In this work, we propose a model design considering inevitable system delays in practical deployment with two design variables: future visibility t_f, the temporal offset between the start of each prediction and the latest available input, and output chunk duration k, the number of frames emitted per call. We train Transformer decoders across a grid of (t_f,k) and show two consistent trade-offs: increasing effective t_f improves coherence by reducing the recency gap, but requires faster inference to stay within the latency budget; increasing k improves throughput but results in degraded accompaniment due to a reduced update rate. Finally, we observe that naive maximum-likelihood streaming training is insufficient for coherent accompaniment where future context is not available, motivating advanced anticipatory and agentic objectives for live jamming.","primarySubject":"Generative Tasks -> real-time considerations","secondarySubject":"Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p9-2-streaming-generation-for","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5RGET1","funFact":"Yusong used to play the timpani in both symphony and wind orchestras.","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_13.html","displayId":"P9-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Real-time considerations","Music and audio synthesis"]},"477":{"uid":"477","title":"SymSkill: Skill Graph for Symbolic Music Understanding","authors":"Junda Wu; Boyang Wang; Xunyi Jiang; Aditya Yadavalli; Ziyu Wang; Qianqi Yan; Hongyi Wen; Julian McAuley","abstract":"Symbolic music understanding requires ordered, score-grounded procedures combining harmonic, melodic, rhythmic, and formal evidence beyond recognising musical terminology. We introduce SymSkill, a skill-augmented framework that distils such procedures from music-theory curricula into a graph-structured reasoning resource for symbolic-music question answering. Unlike prompt-based methods using unstructured chain-of-thought or tool-augmented agents relying on hand-authored MIR functions, SymSkill represents each music-analysis skill as a typed operator over musicological objects, with a micro-protocol, invariant checks, ambiguity conditions, and expected output. A musicological ontology provides a coordinate system for curriculum-based distillation, graph construction, and test-time retrieval. To construct the skill graph, an LLM-driven analyst--critic self-examination process over the Open Music Theory curriculum generates questions exposing reasoning failures under restricted evidence, which are converted into reusable skills and typed relations. At inference, an ontology-aware retriever renders relevant skills as compact protocols for a tool-free LLM reasoner to apply to symbolic score representations. Across symbolic-music benchmarks, SymSkill gains most when retrieved skills are executed as explicit procedures rather than appended as unstructured context, especially on complex tasks.","primarySubject":"MIR data fundamentals -> web mining, and natural language processing","secondarySubject":"Computational musicology -> computational music theory and musicology;Creativity -> tools for artists;MIR data fundamentals -> symbolic music processing","presentation":"","slackChannel":"p9-3-symskill-skill-graph","channelUrl":"https://ismir2026.slack.com/archives/C0C5T8H4U69","funFact":"","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_477.html","displayId":"P9-03","presenterMode":"","materials":["paper"],"keywords":["MIR data fundamentals","Web mining, and natural language processing","Computational musicology","Computational music theory and musicology","Creativity","Tools for artists","Symbolic music processing"]},"128":{"uid":"128","title":"TART: A Modular Tool For Technique-Aware Audio-To-Tablature Guitar Transcription","authors":"Akshaj Gupta; Hwi Joo Park; Andrea Guzman; Shamak Gowda; Samhita Konduri; Jiachen Lian; Robin Netzorg; Gopala Anumanchipalli","abstract":"Automatic Music Transcription (AMT) for guitar remains limited by three challenges: existing systems often fail to capture expressive techniques such as slides, bends, and percussive hits; they often assign notes to incorrect string-fret combinations; and they are typically trained on clean recordings, limiting generalization to noisy real-world audio. To address these challenges, we propose TART, a modular four-stage audio-to-tablature pipeline consisting of (1) an audio-to-MIDI transcription model, (2) an expressive technique classifier, (3) an audio-conditioned T5 encoder-decoder for string-fret assignment, and (4) an automated tablature generator. We evaluate TART in a zero-shot setting on GuitarSet, EGDB, and two augmented benchmarks, Noisy GuitarSet and Noisy EGDB. Averaged across these four benchmarks, TART achieves 81.35% audio-to-MIDI F50 (+6.67 points over the best prior baseline), 71.8% string-fret Tab F1 (+8.5 points over the best prior baseline), and an overall 54.08% end-to-end Tab F1. To our knowledge, TART is the first framework to generate guitar tablature with both fingering and expressive technique annotations directly from guitar audio.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Applications -> music retrieval systems;Creativity -> tools for artists;MIR data fundamentals -> symbolic music processing;Musical features and properties -> expression and performative aspects of music","presentation":"Virtual","slackChannel":"p9-4-tart-a-modular","channelUrl":"https://ismir2026.slack.com/archives/C0C6STH6E64","funFact":"Akshaj is a self-taught guitarist who plays in a band.","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_128.html","displayId":"P9-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Applications","Music retrieval systems","Creativity","Tools for artists","MIR data fundamentals","Symbolic music processing","Musical features and properties","Expression and performative aspects of music"]},"141":{"uid":"141","title":"PEACE: Joint Embeddings of DSP Effects Code and Audio","authors":"David Braun; Adam Finkelstein","abstract":"This paper introduces PEACE, the first joint embedding of audio effect code and output audio. Building on SLAP's multimodal objective, we pair an AFx-Rep audio encoder with two code encoders for Faust, a functional language for audio signal processing. First, we evaluate a fine-tuned T5 transformer over Faust source code. Second, we evaluate a message passing graph neural network over an intermediate representation of the Faust compiler, capturing both topology and UI parameters. We evaluate on audio-to-code retrieval, where masking UI parameters at inference yields embeddings that encode effect chain topology alone. When parameters are visible, the two code encoders tie on retrieval of mixed-length chains but have tradeoffs on single-effect galleries. With parameters fully masked, PEACE recovers ordered chain topology far above chance without the limitations of supervised methods. PEACE outperforms pretrained models on an out-of-distribution reverb retrieval benchmark and can improve frozen audio-only representations. Its dual understanding of topology and parameters lays groundwork for music information retrieval systems that search, generate, and condition on DSP code.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks -> similarity metrics;Applications -> music composition, performance, and production;MIR data fundamentals -> multimodality","presentation":"Virtual","slackChannel":"p9-5-peace-joint-embeddings","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA75QLB","funFact":"","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_141.html","displayId":"P9-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Similarity metrics","Applications","Music composition, performance, and production","MIR data fundamentals","Multimodality","Open Review"]},"182":{"uid":"182","title":"MIDI-LLM: Improving Text-to-MIDI Music Generation via Adapting Large Language Models","authors":"Shih-Lun Wu; Dave Carlton; Ryan Miyakawa; Yoon Kim; Chris Donahue; Cheng-Zhi Anna Huang","abstract":"We present MIDI-LLM, a recipe that improves multitrack text-to-MIDI generation via adapting Large Language Models (LLMs). MIDI-LLM expands an LLM's text vocabulary to include MIDI tokens and employs a two-stage training pipeline: (i) unimodal continued pretraining on music-adjacent text and standalone MIDIs, and (ii) multimodal supervised finetuning on text–MIDI pairs. Our instantiation of MIDI-LLM based on Llama 3.2 (1B) outperforms the recent Text2midi model in both text control and musical quality, and readily integrates with optimized inference ecosystems like vLLM. To align with real-world songwriting workflows, we further finetune our MIDI-LLM on the TheoryTab dataset for text-conditioned lead sheet (i.e., melody + chords) generation and infilling. A comprehensive ablation study validates the synergy between LLM text pretraining, standalone MIDI pretraining, and supervised text-to-MIDI finetuning. Finally, an in-the-wild blind user study conducted in a real-world creative workflow at scale with 58 Hookpad Aria users and 4,002 generated outputs demonstrates that our MIDI-LLM achieves the highest acceptance rate in zero-to-one lead sheet generation over baselines without text control or LLM pretraining, confirming its efficacy in human-AI music co-creation.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Applications -> music composition, performance, and production;Creativity -> human-AI co-creativity;Generative Tasks -> artistically-inspired generative tasks;MIR data fundamentals -> multimodality;MIR data fundamentals -> symbolic music processing","presentation":"Virtual","slackChannel":"p9-6-midi-llm-improving","channelUrl":"https://ismir2026.slack.com/archives/C0C5YC9JRDY","funFact":"","flags":["student","award"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_182.html","displayId":"P9-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Applications","Music composition, performance, and production","Creativity","Human-AI co-creativity","Artistically-inspired generative tasks","MIR data fundamentals","Multimodality","Symbolic music processing","Awards Nominee","Open Review"]},"357":{"uid":"357","title":"Evaluating Music Context Preservation: A Multi-Facet Framework for Music Editing Systems","authors":"Yash Vishe; Eric Xue; Xunyi Jiang; Zachary Novack; Junda Wu; Julian McAuley; Xin Xu","abstract":"Music editing plays a vital role in modern music production, with applications in film, broadcasting, and game development. Recent advances in music editing systems have enabled diverse editing tasks such as timbre transfer, instrument substitution, and genre transformation. However, many existing works overlook evaluating their ability to preserve musical facets that should remain unchanged during editing, which we define as Music Context Preservation (MuseCP). While some studies do consider MuseCP, their evaluation protocols and metrics are not comprehensive. To address this, we introduce the first MuseCP evaluation framework, MuseCPEval, that covers four categories of music facets with fine-grained and well-tailored metrics to capture nuanced changes in music attributes. Objective validation and a human study demonstrate the effectiveness of these metrics. Moreover, the case studies on diverse music editing systems illustrate the practical utility of these metrics as a testbed and diagnostic tool, providing insights into the strengths and limitations of existing systems. We hope our metrics and findings can offer practical guidance for developing more effective and reliable music editing strategies with strong MuseCP capability.","primarySubject":"Evaluation, datasets, and reproducibility -> evaluation metrics","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology","presentation":"Virtual","slackChannel":"p9-7-evaluating-music-context","channelUrl":"https://ismir2026.slack.com/archives/C0C6STTFJL8","funFact":"1. We received 33 survey responses in total: 17 from the Chinese version and 16 from the English version. Among them, 6 Chinese responses and 5 English responses passed our sanity check.\n2. Many frontier LLMs apparently can’t “hear” music very well. We tried to ask Claude help design edits for our human listening test, but many of its suggestions changed only the underlying data representation while producing almost no perceptible difference to human listeners. After trying several models, Gemini might have the strongest synesthetic intuition. For example, when we talked with it about the walls of the Kimbell Art Museum, it somehow led us to the Salk Institute; when we discussed Le Bateau-atelier, the conversation wandered naturally to Vilhelm Hammershøi. Perhaps the interpretability of synesthesia in world models is an interesting research topic. :)\n3. One of our authors, Xin Xu, has worked with classical music and has played the piano for 22 years. She is passionate about research at the intersection of music and NLP. Feel free to reach out at xinxucs@ucsd.edu if you share similar interests!","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_357.html","displayId":"P9-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Evaluation metrics","Evaluation methodology","Open Review"]},"4":{"uid":"4","title":"FusID: Modality-Fused Semantic IDs for Generative Music Recommendation","authors":"Haven Kim; Yupeng Hou; Julian McAuley","abstract":"Generative recommendation systems have achieved significant advances by leveraging semantic IDs to represent items. Recent work increasingly adopts multimodal approaches, and music recommendation—a domain that has historically leveraged rich multimodal information—is one of the representative application areas. However, existing approaches that tokenize each modality independently face three limitations: redundancy across modalities that reduces efficiency, an inability to capture inter-modal interactions that limits item representation, and a disregard for the varying importance of individual modalities. We introduce FusID, a modality-fused semantic ID framework that addresses these limitations through three key components: (i) multimodal fusion that learns unified representations by jointly encoding information across modalities, (ii) representation learning that brings frequently co-occurring item embeddings closer while maintaining distinctiveness and mitigating feature redundancy across sub-embeddings, and (iii) product quantization that converts the fused continuous sub-embeddings into multiple discrete tokens. Evaluated on next-song recommendation task, FusID outperforms the baselines in MRR and Hit@k (k=1,5,10,20) while achieving zero ID conflicts—ensuring each token sequence maps to exactly one song—and mitigates codebook underutilization. The Python implementation of this paper will be released upon acceptance.","primarySubject":"Applications -> music recommendation and playlist generation","secondarySubject":"","presentation":"Virtual","slackChannel":"p9-8-fusid-modality-fused","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7SPYSH","funFact":"This project started as Haven's class project at UCSD.","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_4.html","displayId":"P9-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications","Music recommendation and playlist generation"]},"428":{"uid":"428","title":"MixWeaver: Stem-Aware Automatic DJ Transition Generation via Structured Discrete Optimization","authors":"Xuran Zhou; Xunyi Jiang; Churan Zhi; Junda Wu; Julian McAuley","abstract":"DJ mixing requires coordinating multiple musical attributes, yet existing approaches in music information retrieval lack a unified framework for modeling these factors. In this work, we propose MixWeaver, an interpretable framework for automatic DJ transition that captures key aspects of transition quality, including musical structure and alignment, stem-wise coordination, transition timing, and perceptual consistency. We formulate DJ transition as an optimization problem over stem-wise gain functions, enabling a principled representation of how musical elements are exchanged between two tracks over time. Our method constructs a candidate space by combining stem-specific transition templates and musically aligned windows for four stems, and evaluates each candidate using an interpretable objective based on stem-overlap control, gain smoothness, and total energy stability. A genetic algorithm is used to search this large combinatorial space efficiently. Experiments on more than 2,000 transition pairs show that the full system consistently outperforms candidate-space and scoring ablations in objective evaluation, and a subjective listening study further shows preference over a rule-based baseline. These results highlight the value of interpretable stem-aware design for automatic DJ transition generation.","primarySubject":"Applications","secondarySubject":"Analysis tasks;Generative Tasks;Generative Tasks -> evaluation metrics;Musical features and properties","presentation":"Virtual","slackChannel":"p9-9-mixweaver-stem-aware","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6FD9E3","funFact":"","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_428.html","displayId":"P9-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Applications","Analysis tasks","Generative Tasks","Evaluation metrics","Musical features and properties"]},"74":{"uid":"74","title":"VIOLET: High-Fidelity Violin Synthesis with Techniques and Dynamics","authors":"Baotong Tian; Cynthia Lu; Vincent K.M. Cheung; Ting-Kang Wang; Jonathan Churchill; Zhiyao Duan","abstract":"Neural synthesis for musical instruments has the potential to revolutionize current practices that use concatenative synthesis and a sample library. However, most research focused on piano synthesis and expressive performance generation; little work has been done on continuously articulated instruments like the violin, let alone rendering them with playing techniques and dynamics. We present VIOLET, a latent-diffusion framework for controllable violin synthesis, which uses a Diffusion Transformer (DiT) with rectified flow to synthesize high-fidelity audio from MIDI notes, playing techniques, and continuous dynamics. To train VIOLET, in addition to using a few existing datasets, we curate a new dataset named CSV-TD, which contains 39 h of 48 kHz synthetic audio and time-aligned annotations of MIDI notes, note-level techniques, and continuous dynamics curves. Objective and subjective evaluations show that VIOLET synthesizes violin performances with high technique adherence, accurate pitch and timing alignment, and good dynamics control. It outperforms the current state-of-the-art neural violin synthesis system and approaches a top commercial virtual instrument in terms of technique clarity, naturalness, and dynamics following.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Applications -> music composition, performance, and production;Creativity -> tools for artists;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Musical features and properties -> expression and performative aspects of music","presentation":"In-person","slackChannel":"p9-10-violet-high-fidelity","channelUrl":"https://ismir2026.slack.com/archives/C0C60EZ01U1","funFact":"The first author is the only one who can't play the violin; the last author even performed at the ISMIR 2024 Jam Session!","flags":["student"],"session":"PS-09","posterUrl":"https://ismir2026program.ismir.net/poster_74.html","displayId":"P9-10","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Applications","Music composition, performance, and production","Creativity","Tools for artists","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Musical features and properties","Expression and performative aspects of music"]},"490":{"uid":"490","title":"Pitch Contour Tokenization using VQ-VAE and Its Application on Korean Traditional Music Analysis","authors":"Seonguk Ju; Seola Cho; Suin Chung; Danbinaerin Han; Dasaem Jeong","abstract":"Computational analysis of music often relies on discrete representations, yet many musical traditions are organized around continuous pitch movement that resists segmentation into note-like units. For such traditions, the discrete units that analysis would build on are not given in advance. We address this gap by learning a vocabulary of local pitch-contour patterns directly from unlabeled audio, using a VQ-VAE that quantizes fixed-length contour segments into a finite codebook. To make the learned tokens stable across segmentation positions and small variations in timing and pitch range, we train the model with a reconstruction objective evaluated under the best alignment among a set of candidate temporal and pitch-domain transformations. Applied to Korean traditional music, the learned tokens recover information about expert-defined sigimsae categories without supervision, and in pansori individual tokens align with the two principal modes, Gyemyeonjo and Ujo, supporting their use as units for corpus-level analysis of contour-centric traditions.","primarySubject":"Analysis tasks -> pattern matching and detection","secondarySubject":"Computational musicology -> computational ethnomusicology;Musical features and properties -> expression and performative aspects of music;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p10-1-pitch-contour-tokenization","channelUrl":"https://ismir2026.slack.com/archives/C0C5T8M30JH","funFact":"","flags":["student","award"],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_490.html","displayId":"P10-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Pattern matching and detection","Computational musicology","Computational ethnomusicology","Musical features and properties","Expression and performative aspects of music","Representations of music","Awards Nominee","Open Review"]},"494":{"uid":"494","title":"VioTech: Frame-Level Violin Playing Technique Detection with Multi-Scale Mixture-of-Experts","authors":"Ting-Kang Wang; Vincent K.M. Cheung; Nan-Tien Lai; Li Su","abstract":"Playing technique is crucial for expressive violin performance, yet few violin technique detection models exist. Current systems mainly rely on note-level pipelines that first estimate note boundaries and then classify techniques. This is especially problematic for bowed string instruments such as the violin, where soft onsets and continuous bow-string interaction render note segmentation unreliable. Here, we propose VioTech, a frame-level violin technique detection framework that predicts techniques directly at 10 ms resolution without explicit note segmentation. To handle the large temporal-scale variation among techniques, VioTech combines a frozen MERT encoder with a multi-scale Mixture-of-Experts (MoE) architecture and task-specific routing for tonal, articulation, and legato-related predictions. We also introduce VioTech120, the first manually annotated dataset of violin playing techniques constructed from real solo performance recordings, as well as an accompanying note-labeling user interface. Results demonstrate that VioTech achieved consistent improvements over prior methods and significantly enhanced cross-dataset generalization, establishing a new state-of-the-art in violin technique-aware transcription.","primarySubject":"Analysis tasks -> automatic classification","secondarySubject":"Analysis tasks -> music transcription and annotation;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Musical features and properties -> expression and performative aspects of music;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"Virtual","slackChannel":"p10-2-viotech-frame-level","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJVJJQ6","funFact":"","flags":[],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_494.html","displayId":"P10-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Automatic classification","Music transcription and annotation","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Musical features and properties","Expression and performative aspects of music","Timbre, instrumentation, and singing voice"]},"509":{"uid":"509","title":"STRAdi: Real-Time Violin Transcription for Score Following","authors":"Sein Lee; Jiyun Park; Juhan Nam","abstract":"Automatic music transcription and score following have been extensively studied for piano, but violin remains underexplored. Weak onsets, continuous pitch variation from vibrato and portamento, limited training data, and frequent silences in ensemble settings make both tasks particularly challenging. Moreover, existing violin transcription models have mainly focused on offline processing, which limits their use in real-time applications. To address this gap, we propose STRAdi, a fully causal violin transcription model designed for streaming inference, and pair it with a symbolic tracker for real-time score following. Despite strict causality, STRAdi remains competitive with its offline non-causal counterpart. The resulting transcribe-and-track pipeline improves tracking robustness over audio-based baselines. In particular, the pipeline is more robust to violin-silent passages than audio-based trackers.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Analysis tasks -> alignment, synchronization, and score following;Generative Tasks -> real-time considerations","presentation":"Virtual","slackChannel":"p10-3-stradi-real-time","channelUrl":"https://ismir2026.slack.com/archives/C0C5T8LDLQM","funFact":"","flags":["student"],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_509.html","displayId":"P10-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Alignment, synchronization, and score following","Generative Tasks","Real-time considerations","Open Review"]},"511":{"uid":"511","title":"Measuring Cross-Cultural Style Diffusion Through Era Classification: US and Korean Popular Music","authors":"Dasol Lee; Minhee Lee; Seonguk Ju; Daewoong Kim; Harin Lee; Dasaem Jeong","abstract":"Popular music circulates globally while being locally reinterpreted, yet this process of cross-cultural style diffusion has rarely been quantified. We propose an era-classification framework for measuring temporal alignment between chart cultures. CNN classifiers trained from scratch on Billboard Hot 100 audio are applied to Korean Melon chart songs. Korean chart songs from the 1960s through the 1980s are consistently inferred as belonging to earlier Billboard eras, by a median of about four to five years, while the same models remain unbiased on held-out Billboard audio. The offset then halves at the 1990s, to roughly two to three years, and holds there through the 2000s. Reverse inference shows a complementary narrowing, and the pattern holds across architectures and seeds. We interpret these results as reflecting how globally circulating pop styles were locally adopted and progressively synchronized. The framework can be applied to other pairs of chart cultures beyond the US–Korea case examined here.","primarySubject":"Computational musicology","secondarySubject":"Analysis tasks -> automatic classification;Computational musicology -> systematic musicology;Musical features and properties -> musical style and genre","presentation":"In-person","slackChannel":"p10-4-measuring-cross-cultural","channelUrl":"https://ismir2026.slack.com/archives/C0C627Z3SGL","funFact":"","flags":["student"],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_511.html","displayId":"P10-04","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Computational musicology","Analysis tasks","Automatic classification","Systematic musicology","Musical features and properties","Musical style and genre","Open Review"]},"75":{"uid":"75","title":"DDSynth-RL: Audio Synthesizer Inversion via Discrete Diffusion with Reinforcement Learning","authors":"Tristan WU; Daniel Chin; Junan Zhang; Junyan Jiang; Yansen Jing; Gus Xia","abstract":"Synthesizer inversion is challenging for two main reasons: 1) Distinct parameter configurations can produce perceptually similar sounds. 2) Parameter-space losses often fail to reflect rendered audio similarity, while the synthesizer being a non-differentiable black box prevents simple audio-domain supervision. To address the one-to-many mapping induced by the first challenge, we formulate synthesizer inversion as conditional generation over discrete synthesizer parameters and use masked discrete diffusion as the generator. This treatment additionally avoids the fixed-order assumption of autoregressive models and the continuous-relaxation mismatch of flow matching when modeling categorical synthesizer controls. To address the second challenge, we further fine-tune the model with GRPO-style audio-domain rewards computed from rendered outputs. Experiments on Dexed show that, after supervised training, the discrete diffusion model is competitive with autoregressive and flow-matching baselines, and reward-based fine-tuning further improves out-of-domain audio matching performance. Code and demos are available at: https://github.com/DDSynth-RL/DDSynthRL.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Analysis tasks -> music transcription and annotation;Applications -> music composition, performance, and production;Generative Tasks -> evaluation metrics;MIR data fundamentals -> symbolic music processing;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"In-person","slackChannel":"p10-5-ddsynth-rl-audio","channelUrl":"https://ismir2026.slack.com/archives/C0C6278BKNY","funFact":"1. Tristan's years of making electronic music for fun helped shape DDSynth-RL. Its epsilon top-k strategy was inspired by randomly tweaking synthesizer knobs to find better sounds.\n\n2. Handcrafted presets reflect sound designers' tastes, which acoustic similarity scores may not fully capture.\n\n3. Reward weights reflect how we perceive sound differences, and may need to change with the sound type and context.","flags":["student","award"],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_75.html","displayId":"P10-05","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Analysis tasks","Music transcription and annotation","Applications","Music composition, performance, and production","Evaluation metrics","MIR data fundamentals","Symbolic music processing","Musical features and properties","Timbre, instrumentation, and singing voice","Awards Nominee","Open Review"]},"208":{"uid":"208","title":"Interpretable Inverse Acoustic Modeling of Percussion via Physics-Guided Spectral Decomposition","authors":"Rajeev Rajan; Kevin Benty; Athul Joe Joseph Palliparambil","abstract":"We propose a physics-based parametric framework for percussive instruments that acts as a digital twin, enabling sound synthesis through direct control of physical parameters such as membrane tension, strike position, and geometry. The model generates sound from first principles using eigenmodes of a non-uniform membrane, yielding a physically interpretable representation. We introduce Physics-Guided Spectral Decomposition (PGSD), an inverse method that aligns the model with observed audio via closed-form ridge regression, enabling estimation of modal amplitudes and underlying physical parameters. Experiments on synthetic data demonstrate physically consistent behavior under parameter variation, while evaluation on real Chenda recordings shows accurate reconstruction (coefficient of determination, R2 up to 0.8152) and consistent excitation trends. The proposed framework provides a unified, bidirectional approach for controllable synthesis and interpretable parameter inference without requiring training data.","primarySubject":"Computational musicology","secondarySubject":"","presentation":"","slackChannel":"p10-6-interpretable-inverse-acoustic","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJ6923G","funFact":"1.The Chenda is a South Indian drum you may have heard at temple festivals and in Kathakali performances in Kerala. Our work turns its sound into physical parameters like membrane tension and strike position.\n2.Shape changes how well the method works. An oval-shaped drum shell made PGSD more accurate than cylindrical or hourglass shells, because its asymmetry stops several vibration modes from sharing the same frequency.\n3.It runs fast without any training. PGSD can estimate a drum's tension and where it was struck in under half a second on a laptop, with no training data. The code is at github.com/arcadehero3-eng/CHENDA_PGSD.","flags":[],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_208.html","displayId":"P10-06","presenterMode":"","materials":["paper","poster","video"],"keywords":["Computational musicology","Open Review"]},"246":{"uid":"246","title":"End-to-End Motif Discovery for Symbolic Polyphonic Music via Differentiable Label Propagation on Note-Pair Graphs","authors":"Raynaldi Lalang; Kazuyoshi Yoshii","abstract":"Symbolic polyphonic motif discovery is a nested partitioning problem that aims to find thematically important note sequences that repeat within a musical piece. A previous method attempts to learn motif characteristics at the note level and then performs pattern discovery as an independent point-set search with a fixed similarity rule. As such, the learned note filter and the fixed search are optimized separately, and the geometric search tends to produce more motif groups and occurrences than necessary. To solve this problem, we propose end-to-end differentiable training that uses a note-pair graph to learn affinities between neighboring notes and form motif occurrences. Annotated occurrences provide soft-membership seeds that propagate through the same affinities used at test time, and differentiable losses at the note-pair, occurrence, and motif-type levels train the affinities jointly. On the BPS-motif dataset (first movements of Beethoven's 32 piano sonatas), our method improved establishment F1 from 0.676 to 0.732 and three-layer F1 from 0.273 to 0.286, and cut occurrences per piece from 1181.1 to 105.1 (ground truth 154.5). Adding occurrence completion gave the highest occurrence precision and the lowest cluster impurity, and yielded 133.4 occurrences per piece.","primarySubject":"Musical features and properties -> melody and motives","secondarySubject":"MIR data fundamentals -> symbolic music processing;Musical features and properties -> representation learning;Musical features and properties -> structure, segmentation, and form","presentation":"","slackChannel":"p10-7-end-to-end","channelUrl":"https://ismir2026.slack.com/archives/C0C6STKT8TA","funFact":"The author played the tuba for his high school music ensemble, although he was not very good at it.\nThis is the author's first work in the MIR field; some of the ideas were inspired by his previous works (coastal engineering and e-commerce recommendation systems).\nThe author is (or rather, dreams of becoming) a music producer.","flags":["student"],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_246.html","displayId":"P10-07","presenterMode":"","materials":["paper","poster","video"],"keywords":["Musical features and properties","Melody and motives","MIR data fundamentals","Symbolic music processing","Representation learning","Structure, segmentation, and form","Open Review"]},"287":{"uid":"287","title":"Pianoid: 3D Pianist Body Motion Generation from MIDI","authors":"Joonhyung Bae; Juhan Nam","abstract":"Pianist performance is a full-body activity, with torso sway, shoulder lift, and head tilt conveying expression alongside the notes. Yet computational models of pianist body motion have so far addressed only hands or 2D upper bodies. We present Pianoid, the first 3D 26-joint pianist body motion system from MIDI. Pianoid is a four-stage hybrid pipeline whose body partition and diffusion placement are derived from a coupling analysis of the MOSA performance corpus. MPJPE (Mean Per Joint Position Error) rewards the near-static collapse of mean-regressed baselines. To surface this failure mode we report per-axis expressiveness ratios aggregated as Δ1, validated by a dual perceptual study (non-expert N=28, expert N=9). Non-expert viewers prefer Pianoid over every baseline and ablation, while still preferring the ground truth. Experts attribute the residual gap to fine timing and articulation rather than under-expression. Pianoid establishes note-level pianist body motion as a tractable MIR task via data-derived hierarchical modeling with selective diffusion.","primarySubject":"Generative Tasks -> artistically-inspired generative tasks","secondarySubject":"MIR data fundamentals -> multimodality;Musical features and properties -> expression and performative aspects of music","presentation":"","slackChannel":"p10-8-pianoid-3d-pianist","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCD1ZJ6","funFact":"","flags":["student"],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_287.html","displayId":"P10-08","presenterMode":"","materials":["paper","poster","video"],"keywords":["Generative Tasks","Artistically-inspired generative tasks","MIR data fundamentals","Multimodality","Musical features and properties","Expression and performative aspects of music","Open Review"]},"84":{"uid":"84","title":"StemFX: Learning Mixing Style Representations via Autoregressive FX Chain Prediction on Source-Separated Stems","authors":"Yuan-Chiao Cheng; Jui-Te Wu; Brian Chen; Yen-Tung Yeh; Yu-Hua Chen; Yi-Hsuan Yang","abstract":"Audio mixing style encompasses the artistic and technical decisions a mix engineer makes, including level balancing, spatialization, and the choice, ordering, and parameterization of audio effects (FX) on each stem. FX chains are a key determinant of this style, yet existing approaches to modeling them remain limited. Some operate on stereo mixtures without explicit per-stem FX chain modeling, others fix the number or type of effects per track, and many require differentiable effect implementations or scarce multitrack datasets. We present StemFX, a framework that learns mixing style representations by autoregressively predicting variable-length FX chains on source-separated stems. A Transformer decoder predicts tokenized FX chains autoregressively, while a band-split multi-band CNN encoder with FiLM conditioning captures per-stem spectral structure. To enable large-scale paired training, we extract pseudo-stems from about 105K songs via source separation and augment them using MultiAFx, a toolkit unifying 85 audio effects from 7 Python libraries. Evaluated on mixing style retrieval, StemFX outperforms all baseline models across all tested chain lengths. On paired mixing style transfer, StemFX achieves the best spectral fidelity and the highest listener preference, over 4000 times faster than iterative optimization.","primarySubject":"Applications -> music composition, performance, and production","secondarySubject":"Analysis tasks -> similarity metrics;Applications -> music retrieval systems;Generative Tasks -> transformations;Musical features and properties -> representation learning;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"","slackChannel":"p10-9-stemfx-learning-mixing","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5TG09M","funFact":"","flags":[],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_84.html","displayId":"P10-09","presenterMode":"","materials":["paper","poster","video"],"keywords":["Applications","Music composition, performance, and production","Analysis tasks","Similarity metrics","Music retrieval systems","Generative Tasks","Transformations","Musical features and properties","Representation learning","Timbre, instrumentation, and singing voice"]},"338":{"uid":"338","title":"Mitigating Structural Collapse in Multi-Codebook Autoregressive Music Generation via Structure-Aware Scheduled Sampling","authors":"Chen Feng Tsai; Mi-Yen Yeh","abstract":"Hierarchical autoregressive music models suffer from \"structural collapse\" during long-horizon generation, caused by exposure bias compounding across the two-dimensional residual vector quantization (RVQ) grid. Standard scheduled sampling exacerbates this by creating unrealistic \"chimeric\" frames. To address this, we propose Structure-Aware Scheduled Sampling (SASS), which performs joint token replacement across codebooks using temperature-scaled Top-K sampling and confidence gating. Experiments on MusicGen demonstrate that SASS reduces Kullback-Leibler divergence by 5.6–10.0%, flattens late-stage ΔFAD degradation, and improves beat regularity by up to 14.2%. Listening studies confirm that listeners prefer SASS for coherence and rhythmic stability in 20–30s segments.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology;Generative Tasks -> evaluation metrics;Musical features and properties -> representation learning","presentation":"","slackChannel":"p10-10-mitigating-structural-collapse","channelUrl":"https://ismir2026.slack.com/archives/C0C60F9AG8Z","funFact":"1. I used to work in a music studio that produce anime music.\n2. I went to 2 EDC festivals this year, and going to another one EDC in December in Thailand.\n3. I have stank face and bass face for Jazz and EDM.","flags":[],"session":"PS-10","posterUrl":"https://ismir2026program.ismir.net/poster_338.html","displayId":"P10-10","presenterMode":"","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Evaluation, datasets, and reproducibility","Evaluation methodology","Evaluation metrics","Musical features and properties","Representation learning","Open Review"]},"373":{"uid":"373","title":"Emergent Tonal Structure in Learned Chord Embeddings and Its Relation to Tonal Tension","authors":"Maral Ebrahimzadeh; Gilberto Bernardes; Sebastian Stober","abstract":"Several tonal pitch spaces and computational models have been proposed to analyze tonal structure in Western tonal music, many of them grounded in principles from music theory and used to support tonal analysis with important implications for tonal tension. In parallel, data-driven methods such as skip-gram have been used to learn chord embeddings from symbolic corpora, but their ability to recover tonal structure and its relation to tonal tension remains underexplored. In this work, we investigate how skip-gram chord embeddings reflect tonal structure and whether they provide a useful basis for analyzing structural aspects of tonal tension. Using chord sequences with and without transposition-based augmentation, we evaluate the learned spaces from geometric, functional, and tension-related perspectives. We show that augmented embeddings exhibit strong transposition equivariance, recover a clear circle-of-fifths structure, and support interpretable shifts between key-related regions of the learned space. We then derive embedding-based measures from chord-to-key distance and contextual chord-distance relations, and show that they capture meaningful aspects of tonal tension structure through correspondence with matched tonal measures and moderate alignment with human tension profiles. Across analyses, transposition-based augmentation generally improves the stability, tonal coherence, and interpretability of the learned space.","primarySubject":"Musical features and properties -> harmony, chords and tonality","secondarySubject":"Analysis tasks -> similarity metrics;MIR data fundamentals -> symbolic music processing;Musical features and properties -> representation learning","presentation":"Virtual","slackChannel":"p11-1-emergent-tonal-structure","channelUrl":"https://ismir2026.slack.com/archives/C0C627NTM5J","funFact":"","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_373.html","displayId":"P11-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Harmony, chords and tonality","Analysis tasks","Similarity metrics","MIR data fundamentals","Symbolic music processing","Representation learning","Open Review"]},"379":{"uid":"379","title":"MambaVoice: Lightweight Audiovisual Singing Voice Separation via a Hybrid Mamba-Transformer Model","authors":"Adithi Shankar; Gopika Krishnan; Gloria Haro; Xavier Serra; Martín Rocamora","abstract":"Isolating a target singing voice from a music video remains challenging, particularly in the presence of multiple vocalists and dense instrumental accompaniment. We propose MambaVoice, a lightweight audiovisual framework that leverages a hybrid Mamba–Transformer architecture for targeted singing voice separation. The model jointly encodes audio and visual streams using an attention-based band-split audio encoder and a spatio-temporal graph convolutional network (ST-GCN) for facial motion features. These modalities are fused through a multiplicative gating mechanism, enabling visual cues to selectively modulate audio representations. The fused features are processed by a hybrid backbone that combines Transformer self-attention with Selective State Space Models (SSMs), achieving efficient long-range temporal modeling with linear complexity. We evaluated MambaVoice on the Acappella and URSing datasets under challenging conditions, including mixtures with interfering singers. At 16.2 million parameters, the model demonstrates comparable performance, achieving 14.18 dB SDR on Acappella and strong cross-dataset performance on URSing, comparable to larger models at a fraction of the parameter count. These findings highlight the effectiveness of hybrid SSM–attention architectures for scalable, efficient audiovisual source separation, suggesting they are well-suited as lightweight components within larger pipelines. We conduct a perceptual study that further supports our improvements in objective metrics. We provide our implementation online.","primarySubject":"Analysis tasks -> sound source separation","secondarySubject":"Applications -> music videos, multimodal music systems;MIR data fundamentals -> multimodality","presentation":"Virtual","slackChannel":"p11-2-mambavoice-lightweight-audiovisual","channelUrl":"https://ismir2026.slack.com/archives/C0C5WJGGS1L","funFact":"The Mamba model is called as Mamba becuase the model builds on S4 (Structured State Space Sequence models), but adds a selection mechanism computed via a parallel scan. The authors abbreviated this to S6 because pronouncing six \"S\" sounds in a row (S-S-S-S-S-S) sounds exactly like a snake's hiss.","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_379.html","displayId":"P11-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Sound source separation","Applications","Music videos, multimodal music systems","MIR data fundamentals","Multimodality","Open Review"]},"386":{"uid":"386","title":"AllMusicCaps: Album Reviews as Complementary Supervision for Music CLAP","authors":"Pablo Alonso-Jiménez; Xavier Lizarraga-Seijas; Xavier Serra; Dmitry Bogdanov","abstract":"Recent open text-audio contrastive models (CLAPs) are typically trained with LLM-generated captions derived from tag datasets or web search results, which tend to be accurate but expressively narrow. As a complementary source, we explore human-written album reviews, specifically expert reviews from AllMusic: they exist at scale and carry narrative cues, evaluative adjectives, and scene framing that other sources lack. Since raw reviews are too noisy for direct use as captions, we first build a caption corpus with 245,346 samples via an LLM preprocessing pipeline that identifies descriptive musical quotes and rewrites them into training-ready captions. We find that album review supervision yields the largest retrieval gains on a human-written caption benchmark (Song Describer), particularly for complex queries that other existing caption datasets leave uncovered. In addition, we revisit the training recipe and show that SigReg regularization, which encourages an isotropic Gaussian distribution in the embedding space, improves MLP probing across classification tasks, as well as text-to-music retrieval. The resulting model outperforms open CLAP-style baselines on text-to-music retrieval, zero-shot classification, and most MLP probing tasks. We release the review-derived caption dataset and model weights to support future research.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks -> indexing and querying;Applications -> music retrieval systems;MIR data fundamentals -> lyrics and other textual data;MIR data fundamentals -> metadata, tags, linked data, and semantic web","presentation":"Virtual","slackChannel":"p11-3-allmusiccaps-album-reviews","channelUrl":"https://ismir2026.slack.com/archives/C0C5T88683X","funFact":"","flags":[],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_386.html","displayId":"P11-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Indexing and querying","Applications","Music retrieval systems","MIR data fundamentals","Lyrics and other textual data","Metadata, tags, linked data, and semantic web","Open Review"]},"397":{"uid":"397","title":"Comparing Compression-Based Models of Similarity in Jazz Harmonic Progressions","authors":"Xinyi Guan; Edward T. R. Hall; Zeng Ren; Marina Borsodi-Benson; Claire Pelofi; Martin Rohrmeier","abstract":"The ability to accurately account for musical similarity is essential to a wide range of tasks in MIR and cognition. Compression distance provides a promising and underexplored model for similarity: two pieces are similar to the extent that a compact description of one helps describe the other. This paper investigates two competing compression strategies, focusing on surface and structural regularities, respectively. The first estimates how predictable one sequence is under a probabilistic model trained on another using the Prediction by Partial Matching algorithm. The second approach measures how much a piece's derivation tree can be simplified with the learned abstractions from another, using a library-learning algorithm (Stitch). Given a collection of derivation trees, it treats each instance as a program and compresses the collection by learning a library of reusable program abstractions that capture recurring substructures and rewriting the programs in terms of these abstractions. We apply both models to a corpus of jazz harmony analyses and measure their compression performance, comparing interaction between the two. We further present qualitative analyses to demonstrate where the models diverge and agree.","primarySubject":"Computational musicology -> computational music theory and musicology","secondarySubject":"Analysis tasks -> similarity metrics;Cognitive and user-centered MIR -> perception and cognition;Musical features and properties -> harmony, chords and tonality;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p11-4-comparing-compression-based","channelUrl":"https://ismir2026.slack.com/archives/C0C5H68A7B9","funFact":"","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_397.html","displayId":"P11-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Computational musicology","Computational music theory and musicology","Analysis tasks","Similarity metrics","Cognitive and user-centered MIR","Perception and cognition","Musical features and properties","Harmony, chords and tonality","Representations of music"]},"405":{"uid":"405","title":"RNGen: A Large Language Model Approach to Roman Numeral Analysis Toward Explainable Harmonic Analysis","authors":"Hewei Gao; Tim Beyer; Xiaoxuan Wang; Martin Rohrmeier","abstract":"Roman Numeral Analysis (RNA) is often cast as a classification task, yet it is inherently interpretive: multiple analyses may be musically plausible, and their value often depends not only on the labels themselves but also on the reasoning and analytical perspective that support them. While existing RNA systems achieve strong performance as classification models, they are primarily optimized for label prediction rather than for serving as a foundation for rationale-bearing harmonic analysis. In this work, we present RNGen, a large language model approach to Roman Numeral Analysis in symbolic music, as a first step toward such systems. RNGen bridges a symbolic music encoder and a large language model through a lightweight projector, adapting a vision-language-style architecture to harmonic analysis. To make this bridge effective, we introduce a series of music-specific adaptations spanning representation design, musically informed data processing, and staged multi-task fine-tuning. Experiments show that this encoder-to-LLM approach can model RNA competitively, and ablations confirm the importance of these adaptations. We position RNGen not as a completed solution to explainable RNA, since it does not yet produce validated explanations, but as the first LLM-based foundation for it: a framework that demonstrates the feasibility of LLM-based Roman Numeral Analysis from symbolic music and provides a foundation for future systems that can support analytical conclusions with musically meaningful justifications. The code and data for this work are available at https://github.com/Hurwitzzz/RNGen.","primarySubject":"Musical features and properties -> harmony, chords and tonality","secondarySubject":"Computational musicology -> computational music theory and musicology;MIR data fundamentals -> multimodality;MIR data fundamentals -> symbolic music processing;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p11-5-rngen-a-large","channelUrl":"https://ismir2026.slack.com/archives/C0C5T8964UV","funFact":"1. This was Hewei’s very first Music AI project, making it a personal milestone.\n2. If we’d had coding agents back then, we might have finished this project in just 20% of the time!\n3. Plot twist: one of the authors went on to start a company in Silicon Valley!","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_405.html","displayId":"P11-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Harmony, chords and tonality","Computational musicology","Computational music theory and musicology","MIR data fundamentals","Multimodality","Symbolic music processing","Representations of music","Open Review"]},"406":{"uid":"406","title":"AudioCodex: Streaming Autoencoders for Live Audio","authors":"Marco Pasini; Javier Nistal; Cyran Aouameur; Stefan Lattner; György Fazekas","abstract":"We introduce AudioCodex, a fully causal autoencoder for live generative audio applications that compresses stereo audio into compact continuous latent sequences at ∼10 Hz. It combines a Mean Flow magnitude autoencoder with latent hypersphere projection, bounded-noise augmentation, representation alignment, and adversarial phase recovery to learn representations that support both efficient downstream generation and high-quality reconstruction. Through extensive ablations, we show that AudioCodex outperforms the state-of-the-art in terms of generation-reconstruction trade-off, whilst remaining strictly causal in both encoding and decoding. This work paves the way towards truly real-time interactive and streaming generative audio applications. Supplementary results and audio samples are available in the website.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Generative Tasks -> real-time considerations;Musical features and properties -> representations of music","presentation":"Virtual","slackChannel":"p11-6-audiocodex-streaming-autoencoders","channelUrl":"https://ismir2026.slack.com/archives/C0C6STXDNV6","funFact":"","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_406.html","displayId":"P11-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Real-time considerations","Musical features and properties","Representations of music","Open Review"]},"87":{"uid":"87","title":"Improved Robustness in AI-Generated Music Detection","authors":"Emile Dugelay; Thomas Barand; Aurélien Laouar; Baptiste Campeas; Darius Afchar; Romain Hennequin","abstract":"AI music generators leave predictable spectral artifacts determined by their architecture. Existing detectors exploit these artifacts with near-perfect accuracy on raw generated tracks, but their performance collapses under simple audio manipulations, such as speed modification or pitch shifting. We address this open robustness problem by introducing a frequency-scaling-invariant detection pipeline that aims to prevent this kind of attack by design. Our method maps audio onto a log-frequency axis via a log-STFT remapping. A single learned cross-correlation filter, combined with max-pooling, provides shift invariance at inference time. Training uses a hybrid loss that jointly supervises binary detection and artifact-peak localization, regularizing boundary weights. Because robustness to speed change is built in by design, the detector is also interpretable: it outputs both a binary decision and an estimate of the applied speed-change factor.","primarySubject":"Analysis tasks -> automatic classification","secondarySubject":"Analysis tasks -> pattern matching and detection;MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p11-7-improved-robustness-in","channelUrl":"https://ismir2026.slack.com/archives/C0C5SA5LGDR","funFact":"","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_87.html","displayId":"P11-07","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Automatic classification","Pattern matching and detection","MIR data fundamentals","Music signal processing","Open Review"]},"289":{"uid":"289","title":"\"More than words\": Genre-Discriminating Lyrical Patterns in Anglophone German Popular Music","authors":"Markus Radke; Steffen Lepa","abstract":"Recent advances in natural language processing have improved lyrics-based music genre classification through large language models, yet such black box systems obscure the actual linguistic style differences between music genres. Interpretable approaches in computational musicology and MIR, meanwhile, often relied on either formal structural analysis of lyrics or on topic or sentiment analyses. However, genre-typical formulaic expressions, though deemed relevant in popular music studies, have hardly been investigated at all. Our study contributes to filling this gap by uncovering genre discriminative lyrical style patterns—topics, sentiments, and formulaic expressions—in fifty years of Anglophone German popular music. Using the POPTRAG corpus and Discogs genre annotations for 126,042 tracks with English lyrics, we apply the fightin' words method to identify statistically distinctive n-grams across genres. We then identify latent topics, sentiments, and types of formulaic expressions from these via structural topic modeling using training corpus target genres as prevalence covariates. Resulting interpretable lyrics features are employed in machine-learning models that reveal the function of stylistic markers for specific popular music genres. Our transparent ML method approaches the predictive performance of a transfer-learned DistilBERT classifier while offering conceptual insight into how lyrics encode genre identity.","primarySubject":"Computational musicology","secondarySubject":"Analysis tasks -> automatic classification;Computational musicology -> systematic musicology;MIR data fundamentals -> lyrics and other textual data;MIR data fundamentals -> web mining, and natural language processing;Musical features and properties -> musical style and genre","presentation":"Virtual","slackChannel":"p11-8-more-than-words","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7Z6SUD","funFact":"- Markus Radke is a former recording producer for the Singapore Symphony Orchestra who turned to popular music for his PhD and lyrics for this research.\n- Steffen Lepa originally studied communication science and now returns back to his disciplinary roots with this work.","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_289.html","displayId":"P11-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Computational musicology","Analysis tasks","Automatic classification","Systematic musicology","MIR data fundamentals","Lyrics and other textual data","Web mining, and natural language processing","Musical features and properties","Musical style and genre","Open Review"]},"206":{"uid":"206","title":"Recovering Affective Signals in Baroque Opera: A Comparative Evaluation of Human and LLM Annotation","authors":"Carlos Vaquero Patricio; Álvaro Torrente","abstract":"We investigate whether affective annotations of 18th-century opera aria texts can be predicted from symbolic musical features beyond chance-level. Leveraging the Metastasian repertoire, in which single libretti were set by dozens of composers, we analyse 1,668 distinct musical settings across 85 aria texts. Evaluating 19 conditions, we contrast human expert annotations against zero-shot LLM labelling (Gemma 3, Phi-4) over three label space types: human-defined, historically motivated (Descartes' taxonomy of passions), and LLM-proposed. Classification skill is assessed using symbolic music features under aria-grouped cross-validation, normalised against a stratified dummy baseline. While human schemes yield statistically significant but modest predictive signal (skill=0.024-0.045), Gemma 3 re-annotation more than doubles this performance. Notably, the best-performing conditions across historically motivated and LLM-proposed taxonomies achieve skill comparable to human schemes, suggesting that composers' choices reflect textual affect under multiple taxonomy designs with capable annotation. Negative silhouette coefficients across all conditions indicate that affective categories do not form geometrically coherent clusters, suggesting that Baroque composers realised the same affect through musically heterogeneous strategies. However, the significant recovery of text-derived labels indicates that affective distinctions encoded in Metastasio's aria texts remain detectable across diverse musical settings despite compositional variance.","primarySubject":"MIR data fundamentals -> lyrics and other textual data","secondarySubject":"Applications -> digital libraries and archives;Computational musicology -> systematic musicology;Evaluation, datasets, and reproducibility -> evaluation methodology;MIR data fundamentals -> symbolic music processing;Musical features and properties -> musical affect, emotion and mood","presentation":"","slackChannel":"p11-9-recovering-affective-signals","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5WPVC7","funFact":"","flags":[],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_206.html","displayId":"P11-09","presenterMode":"","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Lyrics and other textual data","Applications","Digital libraries and archives","Computational musicology","Systematic musicology","Evaluation, datasets, and reproducibility","Evaluation methodology","Symbolic music processing","Musical features and properties","Musical affect, emotion and mood"]},"309":{"uid":"309","title":"Human Performance in Lyric Topic Classification: Evaluation and Implications","authors":"Varvara Papazoglou; Robert Gaizauskas; Alexandros Stamatiadis","abstract":"Automatic systems for lyric topic classification typically rely on single-label gold annotations derived from curated external sources, but it remains unclear how well these labels reflect listener interpretations or how well listeners agree with each other on song topics. In this paper, we present, to our knowledge, the first systematic study of human performance on lyric-based topic classification and compare it with a logistic regression baseline on the same dataset. We conduct an online annotation study in which participants assign three ranked topic labels to songs based solely on their lyrics. We evaluate inter-annotator agreement, as well as individual and aggregated human accuracy against Songfacts-derived ground truth labels.\n\nThe results show that lyric topic classification is challenging and inherently subjective. Specifically, individual annotators match the gold label 28% of the time (53% within their top three), comparable to an automatic baseline (27%/48%); aggregated human judgements reach 42%/69%, exceeding the strongest automatic configuration on the same songs in top-3 accuracy (33%/63%). Inter-annotator agreement is low to moderate (pairwise τ = 0.29–0.38) but rises consistently when measured against group consensus (τ = 0.40–0.51), indicating that annotators diverge in the details within a partially shared interpretation. These findings suggest that single-label evaluation may underestimate valid human judgements and motivate richer annotation and evaluation schemes for the task.","primarySubject":"Cognitive and user-centered MIR -> user-centered evaluation","secondarySubject":"Analysis tasks -> automatic classification;Cognitive and user-centered MIR;Evaluation, datasets, and reproducibility;MIR data fundamentals -> lyrics and other textual data","presentation":"Virtual","slackChannel":"p11-10-human-performance-in","channelUrl":"https://ismir2026.slack.com/archives/C0C5SAAJ2HZ","funFact":"","flags":["student"],"session":"PS-11","posterUrl":"https://ismir2026program.ismir.net/poster_309.html","displayId":"P11-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","User-centered evaluation","Analysis tasks","Automatic classification","Evaluation, datasets, and reproducibility","MIR data fundamentals","Lyrics and other textual data"]},"413":{"uid":"413","title":"Learning to Predict Performance-induced Emotion Differences in Classical Piano Music","authors":"Joann Ching; Gerhard Widmer","abstract":"Music is often used as a medium for communicating emotion, with performers shaping perceived affect through interpretation. This study addresses the challenge of identifying and predicting subtle changes in perceived emotion that are exclusively due to differences in performance. We focus on classical solo piano music, using a set of 6 commercial recordings of Bach’s Well-Tempered Clavier Book I, annotated in terms of valence and arousal. By encoding the recordings through performance-specific features only, we isolate performance information from aspects of the composition itself, which tend to dominate the overall perceived emotional category. A preliminary analysis validates that these features vary meaningfully across performers. We then propose a relative regression framework, Delta-VA, to predict deviations in valence-arousal relative to an “average” performance, thereby focusing on the changes in emotion brought about by a specific way of playing a piece. In addition to the standard R2 regression score, we introduce geometric evaluation metrics to assess the preservation of pairwise differences between performances. Results indicate high directional consistency with the ground truth, but also a compression in prediction magnitude, indicating that the model tends to underestimate expressive performance effects.","primarySubject":"Musical features and properties -> musical affect, emotion and mood","secondarySubject":"Musical features and properties -> expression and performative aspects of music","presentation":"Virtual","slackChannel":"p12-1-learning-to-predict","channelUrl":"https://ismir2026.slack.com/archives/C0C5SAMCPUK","funFact":"1. Research piano performance — openly rooting for the violin. Obviously, I’m on the other team.\n2. I am good at getting distracted by things that absolutely did not need my attention.\n3. Um..... can't think of a third one... I guess I like music?","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_413.html","displayId":"P12-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Musical affect, emotion and mood","Expression and performative aspects of music"]},"421":{"uid":"421","title":"Diff2Mix: Controllable Music Mixing via Diffusion Models and Differentiable Audio Effects","authors":"Yisu Zong; Jinjie Shi; Joshua Reiss","abstract":"Automatic music mixing aims to combine multitrack audio into a balanced and coherent musical piece. Because the content of different songs and the subjective preferences of mixing engineers jointly shape the final outcome, a practical system should deliver well-balanced mixes while allowing for controllable stylistic variation. However, most existing methods treat automatic mixing and mixing style control as separate tasks, making it difficult for a single system to produce high-quality mixes while remaining editable and style-aware. To address this limitation, this paper presents Diff2Mix, a generative automatic mixing system based on diffusion models and a differentiable mixing console. This system offers two levels of optional user control: a reference song enables overall production style control, and the differentiable mixing console provides explicit audio effects parameters for interpretability and fine-grained optimization. We demonstrate our system’s competitive performance through both objective and subjective evaluations in terms of mixing quality and control ability.","primarySubject":"Applications -> music composition, performance, and production","secondarySubject":"Creativity -> tools for artists;MIR data fundamentals -> music signal processing","presentation":"In-person","slackChannel":"p12-2-diff2mix-controllable-music","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6BNPST","funFact":"","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_421.html","displayId":"P12-02","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Applications","Music composition, performance, and production","Creativity","Tools for artists","MIR data fundamentals","Music signal processing"]},"422":{"uid":"422","title":"Cyclic Attractors in Iterated LLM-Based Symbolic Music Style Transfer","authors":"Jiarui Li; Mark Liberman","abstract":"Large Language Models (LLMs) are increasingly employed as zero-shot tools for symbolic music style transfer. As they become embedded in iterative composition workflows, their behavior under repeated application, rather than single-shot quality, becomes an important but unstudied question. Liberman observed that modern neural text translation systems can converge to a fixed point under round-trip cycling rather than degrading, hinting that similar dynamics may govern other translation tasks. We fed MIDI as compact text into two LLMs and ran a suite of three probing experiments spanning close-pair, three-style, and cross-cultural cycles across Western Classical, Western Pop/Folk, Chinese Pop, British Folk, and Chinese Classical corpora, first verifying that a single application of the operator performs instruction-driven transfer. We find that iterative LLM style transfer converges to a stable attractor in CLaMP2-based style geometry. The attractor's location depends on the cycle's starting style, and in cross-cultural cycles this produces a clear asymmetry: the final endpoint shifts are consistent with model-family-specific style priors. We support these findings with Fréchet Music Distance, a structured 42-listener study, and a format-matched round-trip baseline that separates model behavior from information loss in the text encoding. This extends the cyclic-attractor perspective from language to music, offering a diagnostic route for identifying culturally asymmetric generation in LLM-based creative tools.","primarySubject":"Generative Tasks -> transformations","secondarySubject":"Computational musicology -> computational ethnomusicology;Generative Tasks -> evaluation metrics;Generative Tasks -> qualitative evaluations;MIR data fundamentals -> symbolic music processing;Responsibility and Ethics in MIR -> legal and societal aspects of MIR","presentation":"Virtual","slackChannel":"p12-3-cyclic-attractors-in","channelUrl":"https://ismir2026.slack.com/archives/C0C5YCT5F98","funFact":"Fun Fact 1: I’m a hardcore Hollow Knight / Silksong fan, and collaborating with Christopher Larkin someday is definitely on my bucket list. In fact, this paper was partly born from wanting to hear my favorite game music in completely different styles.  (Also, when are we getting Sea of Sorrow in Silksong?)\n\nFun Fact 2: Some LLM-generated music from our experiments sounded less like “style transfer” and more like my cat sprinting across the piano at 3 a.m.","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_422.html","displayId":"P12-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Transformations","Computational musicology","Computational ethnomusicology","Evaluation metrics","Qualitative evaluations","MIR data fundamentals","Symbolic music processing","Responsibility and Ethics in MIR","Legal and societal aspects of MIR"]},"493":{"uid":"493","title":"Inferring Implied Harmonies through Voice Leading in Monophonic Instrumental Music Using Graph-Based Methods","authors":"Marina Borsodi-Benson; Martin Rohrmeier","abstract":"Latent polyphony is a central feature of Western tonal music, where non-adjacent notes can be understood as belonging to a single voice. Even in a monophonic texture, separate voices can seem to emerge from the musical surface, creating implied harmonies. Modeling latent polyphony not only involves inferring these harmonies, but also capturing the voice leading connections. Existing computational models of music often rely on explicit voices and simultaneously sounding-harmonies to be studied.\n\nWe propose a rule-based method of identifying the implied polyphony of monophonic music and demonstrate how these implied lines can be used to produce a harmonic reduction of the piece. Using a graph-based approach, implied voice lines are actualized, allowing voice leading patterns and dissonances to be identified. Using these, a reduction and alignment process is applied that results in a final harmonic reduction.\n\nResults show that the proposed method achieves performance comparable to existing models of harmonic analysis. Importantly, this approach also requires no training data and runs efficiently, enabling fast analysis and easy reconfiguration. We also present step-by-step qualitative examples that visualize intermediate states and clarify how voice-leading inference and reduction decisions shape the final output.","primarySubject":"Musical features and properties -> harmony, chords and tonality","secondarySubject":"Computational musicology -> computational music theory and musicology;Musical features and properties -> melody and motives","presentation":"In-person","slackChannel":"p12-4-inferring-implied-harmonies","channelUrl":"https://ismir2026.slack.com/archives/C0C5YD06E6N","funFact":"","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_493.html","displayId":"P12-04","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Harmony, chords and tonality","Computational musicology","Computational music theory and musicology","Melody and motives"]},"520":{"uid":"520","title":"Measure for Measure: Reliable Evaluation and a New State of the Art for Full-Page Optical Music Recognition","authors":"Tobias Hornbogen; Josef Valentin","abstract":"The current frontier of Optical Music Recognition (OMR), captured by the Sheet Music Benchmark (SMB), is the full-page transcription of multi-voice scanned piano and ensemble scores. Progress on this frontier is limited by evaluation as much as by models. Established metrics rely on token-level or parser-based comparisons that fail to handle the specific spine structure of **kern and produce scores that are inflated, incomparable across samples, or uncomputable on a large fraction of polyphonic pages. We introduce the Musical Correction Rate (MCR), a format-native metric that parses **kern, ABC, and MusicXML directly, aligns predictions and references at the measure level, and scores individual notes within each measure. Each reported error is localised to a specific note, and in a human study of 46 samples, MCR tracks estimated correction effort more closely than existing metrics. We also train a 184M-parameter DINOv2-based page-level model on a rebalanced and heavily randomised dataset rendered from PDMX. Zero-shot, it reaches 6.38% MCR on SMB and 9.91% on OpenScore Quartets camera pages, versus the larger current state of the art’s 35.92% and 56.23%, respectively, with one-fifth as many parameters.","primarySubject":"Analysis tasks -> optical music recognition","secondarySubject":"Analysis tasks -> similarity metrics;Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> evaluation metrics","presentation":"Virtual","slackChannel":"p12-5-measure-for-measure","channelUrl":"https://ismir2026.slack.com/archives/C0C5H6QCY6B","funFact":"I got interested in OMR because I needed it for my own musical projects and felt like it could work better with SOTA technology","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_520.html","displayId":"P12-05","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Optical music recognition","Similarity metrics","Evaluation, datasets, and reproducibility","Evaluation methodology","Evaluation metrics","Open Review"]},"270":{"uid":"270","title":"What Makes a Good Layer? Assessing the Layer-Wise Intrinsic Properties of Music Foundation Models","authors":"Angelos-Nikolaos Kanatas; Yuexuan Kong; Pablo Alonso-Jiménez; Xavier Serra; Dmitry Bogdanov","abstract":"Music foundation models are commonly used as frozen audio feature extractors, yet selecting which layer to extract from remains largely heuristic. Current practice defaults to fixed depths or multi-layer fusion, with limited understanding of why certain layers transfer better across downstream tasks or how representation quality varies with depth and pre-training paradigm. We conduct a systematic layer-wise analysis of 12 music foundation models spanning three pre-training paradigms (masked modeling, autoregressive modeling, and contrastive learning), characterizing their hidden representations through intrinsic geometric and transformation-based properties. Correlating label-free representation-quality metrics with layer-wise performance across 15 downstream tasks, we find that several metrics track layer quality for genre classification, emotion recognition, automatic tagging, and beat tracking, albeit with varying strength across tasks and pre-training paradigms. However, all metrics fail on tonal tasks such as key estimation and chord recognition, indicating that no single property serves as a general proxy for representation quality across music information retrieval tasks. To address this gap, we introduce a pitch-transposition equivariance measure that captures properties missed by these standard metrics, providing a consistent indicator of tonal quality across model families. Finally, we show that intrinsic metrics can serve as effective proxies for layer selection, matching or outperforming trainable multi-layer fusion methods, particularly in limited-data settings.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks;Analysis tasks -> automatic classification;MIR data fundamentals -> music signal processing;Musical features and properties","presentation":"Virtual","slackChannel":"p12-6-what-makes-a","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7YRQ3X","funFact":"- This is Angelos' first PhD paper at the Music Technology Group (UPF). You might recognize him from ISMIR 2025 in Daejeon, where he presented his very first ISMIR paper.\n- This collaboration started while Yuexuan was a visiting researcher at MTG in Barcelona.\n- Which layer should you use? Often not the last one! Visit https://angeloskanatas.github.io/music-fms-layer-eval/ to learn more. Want to check your own models? Our toolkit implements a bunch of label-free metrics from the literature, for layer analysis or for monitoring them during training: pip install req-metrics","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_270.html","displayId":"P12-06","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Automatic classification","MIR data fundamentals","Music signal processing","Open Review"]},"322":{"uid":"322","title":"A Cross-Cultural Dataset of Instrumental Difficulty for Woodwinds and Strings","authors":"Rui Yang; Mathieu Giraud; Florence Levé","abstract":"The difficulty of a musical score depends partly on the instrument on which a musician plays it. Estimating the instrumental difficulty of scores aims at understanding and modelling the challenges that musicians face when learning or performing music. Instrumental difficulty labels are not universal: For a given instrument, the same musical excerpt may be assigned different grades depending on grading traditions and pedagogical systems. Do these system have different ways of assessing difficulty?\n\nWe introduce CCE-ID, a cross-cultural dataset on instrument difficulty in different pedagogical systems. The dataset contains 336 short score excerpts of Western music, with images and encodings, for eight woodwind and string instruments (flute, oboe, clarinet, bassoon, violin, viola, cello, and double bass) from 23 public-domain pedagogical methods published between 1787 and 1951. The excerpts are annotated by 104 teachers according to their perception of the instrumental difficulty.\n\nThe dataset enables within-system and cross-system difficulty estimation and initial steps towards modelling instrumental difficulty. We present the dataset and analyse agreement patterns across annotators, instruments, and pedagogical systems, and evaluate baseline models linking symbolic score features to perceived difficulty.","primarySubject":"Applications -> music training and education","secondarySubject":"Computational musicology -> digital musicology;Evaluation, datasets, and reproducibility -> annotation protocols;Evaluation, datasets, and reproducibility -> novel datasets and use cases;MIR data fundamentals -> symbolic music processing;Musical features and properties -> melody and motives","presentation":"","slackChannel":"p12-7-a-cross-cultural","channelUrl":"https://ismir2026.slack.com/archives/C0C60F93T97","funFact":"🎣 Apparently, Die Forelle likes following me around. It has now appeared in two of my papers.\n👀 Spot the difference between the Chinese school textbook version (last paper: https://hal.science/hal-05029666v1) and one source adaptation used here (https://gitlab.com/algomus.fr/cce-instrumental-difficulty/-/blob/main/scores/flute/Flute-501.jpg?ref\\_type=heads). Hint: don’t just look at the Jianpu — check the staff notation too :)\n\n\n🧑‍🏫 Recruiting teachers became a small social-media survival game. 🎮\nRejections and silence were basically part of the workflow 😅. One community wanted $48 to join; I turned my RedNote lurker account 🤿 into a research account; Reddit made me space out posts so I wouldn’t look like spam 🤖...still learning how to convince the internet that I’m a real person!","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_322.html","displayId":"P12-07","presenterMode":"","materials":["paper","poster","video"],"keywords":["Applications","Music training and education","Computational musicology","Digital musicology","Evaluation, datasets, and reproducibility","Annotation protocols","Novel datasets and use cases","MIR data fundamentals","Symbolic music processing","Musical features and properties","Melody and motives","Open Review"]},"327":{"uid":"327","title":"Detection of AI-Generated Stems Within Hybrid Human-AI Music","authors":"François Rigaud; Gabriel Meseguer Brocal; Benjamin Martin; Romain Hennequin","abstract":"This paper presents, to the best of our knowledge, the first study on detecting human-AI hybrid music tracks created by mixing human-produced and AI-generated stems. \nBuilding on recent work showing that AI music detectors can identify decoder-related artifacts in fully generated music, we investigate whether such artifacts remain detectable at the stem level after mixing.\nUsing MUSDB18-HQ database in a two-stem vocal + accompaniment setting, we simulate hybrid mixtures by autoencoding individual stems with a neural codec. \nWe compare two strategies combining AI-generated mix detection and source separation. A naive sequential pipeline, where source separation is followed by detection on separated sources, confirms that artifacts associated with an AI-generated stem are not reliably recovered by generic source separation systems. We therefore propose a parallel architecture in which source separation is only used to estimate source-relative energy within the mixture. We then train simple stem-specific binary classifiers that take as input the generated mix prediction together with the relative energy of the target stem on short audio chunks. Averaging chunk-level predictions yields encouraging track-level results, highlighting the potential of such approaches for detecting AI-generated stems in hybrid music.","primarySubject":"Analysis tasks -> automatic classification","secondarySubject":"Analysis tasks -> pattern matching and detection;Analysis tasks -> sound source separation;MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p12-8-detection-of-ai","channelUrl":"https://ismir2026.slack.com/archives/C0C627FBGGL","funFact":"","flags":[],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_327.html","displayId":"P12-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Automatic classification","Pattern matching and detection","Sound source separation","MIR data fundamentals","Music signal processing","Open Review"]},"345":{"uid":"345","title":"Ghost In The Encoder: Decodable Artist Identity Representations In Lyrics-To-Song Generation","authors":"Arhan Vohra; Choenden Kyirong; Laura Ibañez; Martín Rocamora","abstract":"Text-to-song generation models can be prompted to imitate specific artists or regurgitate entire songs from their training data. Although these phenomena have been documented behaviorally on small datasets, little is known about the internal representations that may give rise to them. Prior interpretability work on generative audio has focused on locating semantic concepts such as genre or time signature within model activations. In this work, we show that a trained model can be probed for linearly decodable representations of artist identity from song lyrics alone, without any additional identifiers. Through a controlled case study of ACE-Step 1.5 spanning 2,000 songs across 100 artists, we demonstrate that the artist associated with a given set of lyrics can be identified within the model’s internal activations, and that this conditioning signal propagates from the lyric encoder to the diffusion backbone during inference. These findings indicate that lyrics constitute an artist-level conditioning channel not addressed by prompt-side replication safeguards. More broadly, our work highlights how latent-space analysis can be used to audit what generative music models have implicitly learned from their training data.","primarySubject":"Generative Tasks","secondarySubject":"Analysis tasks;MIR data fundamentals -> lyrics and other textual data;Musical features and properties -> musical style and genre;Musical features and properties -> representation learning;Responsibility and Ethics in MIR -> legal and societal aspects of MIR","presentation":"Virtual","slackChannel":"p12-9-ghost-in-the","channelUrl":"https://ismir2026.slack.com/archives/C0C5T83S8FP","funFact":"","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_345.html","displayId":"P12-09","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Analysis tasks","MIR data fundamentals","Lyrics and other textual data","Musical features and properties","Musical style and genre","Representation learning","Responsibility and Ethics in MIR","Legal and societal aspects of MIR","Open Review"]},"402":{"uid":"402","title":"Exploring Musical Taste : Analyzing User Content Preferences through Attention Mechanisms","authors":"Yan-Martin Tamm; Anna Aljanaki","abstract":"Listener preferences are influenced by multiple factors including sound (timbre, rhythm, harmony), text (lyrics and their emotional and social implications), and social context (popularity, genre identity, the image of the artist). We investigate these facets using corpus analysis methodology on the Music4All-Onion dataset, treating Attentional Factorization Machines (AFMs) as an analytic instrument: the attention weights learned in the course of fitting listening behavior are extracted and aggregated to quantify, for each user, how much each modality contributes to their listening choices. We find that lyric content receives higher mean attention than audio (0.34 vs. 0.28) across users and is the dominant content axis for 79% of users. We also show that per-user attention varies systematically with users' dominant listening genre: among users who mostly listen to Hip-hop, Pop, and RnB, lyric attention is highest, while users preferring electronic music are the group for whom audio attention exceeds lyric. This expected result shows that AFMs provide a viable analytic instrument for individual preference analysis. We also demonstrate how the quality of the underlying representation can limit the analytical strength.","primarySubject":"Cognitive and user-centered MIR -> personalization","secondarySubject":"Applications -> music recommendation and playlist generation;MIR data fundamentals -> lyrics and other textual data;MIR data fundamentals -> multimodality","presentation":"","slackChannel":"p12-10-exploring-musical-taste","channelUrl":"https://ismir2026.slack.com/archives/C0C627PBZT6","funFact":"This paper is the last chapter of Yan-Martin Tamm's PhD thesis. Check out the rest http://tiny.cc/i8ab101! He's graduating this year, so if you have interesting work like this in-house, let him know!","flags":["student"],"session":"PS-12","posterUrl":"https://ismir2026program.ismir.net/poster_402.html","displayId":"P12-10","presenterMode":"","materials":["paper","poster","video"],"keywords":["Cognitive and user-centered MIR","Personalization","Applications","Music recommendation and playlist generation","MIR data fundamentals","Lyrics and other textual data","Multimodality","Open Review"]},"512":{"uid":"512","title":"Separate-and-Detect: Unified Drum Transcription and Stem Generation via Latent Diffusion","authors":"Wei-Han Hsu; Chih-Cheng Chang; Bo-Yu Chen; Li Su; Yi-Hsuan Yang","abstract":"Automatic Drum Transcription (ADT) is commonly formulated as a direct mapping from a music mixture to symbolic drum events. While effective for transcription, this formulation discards the acoustic stems that are useful for editing, remixing, and production. We revisit an alternative separate-and-detect formulation, where a drum source separation front end first produces five editable drum stems, and a fixed onset detector then converts each stem into symbolic events. The separator is built on a five-stem latent diffusion model that jointly generates kick, snare, toms, hi-hats, and cymbals in a compact VAE latent space. We further study two training-only auxiliary branches — an onset branch (OB) and a timbre branch (TB)—which shape the separator during learning but are discarded at inference. Trained on synthetic drum multitracks and evaluated on MDB Drums and ENST-Drums, the proposed pipeline consistently improves over a strong U-Net-based drum separation baseline in overall transcription F1. It also outperforms a representative end-to-end ADT system on kick and snare F1 under our evaluation protocol, while additionally providing separated audio stems. The ablation results show that OB gives the most stable transcription gains, whereas TB changes the trade-off between reconstruction, acoustic stem quality, and onset detection. These results suggest that generative drum demixing can serve not only as a source separation model, but also as a practical front end for interpretable drum transcription.","primarySubject":"Analysis tasks -> sound source separation","secondarySubject":"Analysis tasks -> music transcription and annotation;Creativity -> tools for artists;Generative Tasks -> music and audio synthesis","presentation":"In-person","slackChannel":"p13-1-separate-and-detect","channelUrl":"https://ismir2026.slack.com/archives/C0C60FSD8LR","funFact":"Wei-Han dreamed of becoming an athlete as a child, but eventually found his way into music and AI research.","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_512.html","displayId":"P13-01","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Analysis tasks","Sound source separation","Music transcription and annotation","Creativity","Tools for artists","Generative Tasks","Music and audio synthesis"]},"39":{"uid":"39","title":"Investigating Ethiopian Kiñit in Secular Music: A Self-Supervised Learning and Computational Analysis","authors":"Berihu Yohannes Tesfay; Cheng-Hsin Hsu; Samuel Redie Fishatsion; Li Su","abstract":"Ethiopian secular music is commonly described by a four-kiñit (i.e. mode) framework—Ambassel, Anchihoye, Bati, and Tizita, along with their variations—based on fixed pentatonic scales. However, this framework remains controversial due to inconsistencies in scalar definitions, limited alignment with performance practice, and the presence of microtonal behavior. These challenges motivate a data-driven investigation of the tonal and melodic structure of the kiñit system and the validity of the four-mode theory in real-world performances. We propose an explainable self-supervised learning (SSL) framework that integrates Grad-CAM with skip-gram based melodic pattern mining for music analysis. We also introduce EMIR-extended, a curated dataset of 1,429 recordings (approx. 93.6 hrs) with expert-validated annotations and enriched metadata. Our framework identifies the melodic pitch sequences most strongly associated with specific kiñit categories. Our results reveal a nuanced interaction between fixed modal structures and possibly shifting tonal centers. Interval patterns that appear out of scale become in-scale under pitch shifting, indicating tonic mobility within performances. Furthermore, cent deviation analysis reveals occasional microtonal behavior with high variance and a gradual decrease in microtonal variation over time.","primarySubject":"Computational musicology","secondarySubject":"Computational musicology -> computational ethnomusicology;Computational musicology -> computational music theory and musicology","presentation":"In-person","slackChannel":"p13-2-investigating-ethiopian-ki","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7SGR6H","funFact":"Did you know? Although Ethiopian Kiñit feels ancient, its popular \"four-mode framework\" (Tizita, Bati, Ambassel, Anchihoye) was actually mentioned in the 1960s.\n\nDid you know? We only gave song audio and mode labels to the model—no music theory rules—and Grad-CAM revealed that the model identified intervals associated with the kinit system.","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_39.html","displayId":"P13-02","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Computational musicology","Computational ethnomusicology","Computational music theory and musicology","Open Review"]},"60":{"uid":"60","title":"A Dataset and Benchmark for Optical Music Recognition of String Quartet Scores","authors":"Dongmin Kim; Brian Liu; Jose J. Valero-Mas; Dasaem Jeong","abstract":"Optical music recognition (OMR) transcribes music scores into digital formats. While the field has advanced significantly on monophonic and piano-form scores, multi-part score transcription remains underexplored, largely due to the absence of a suitable dataset. We introduce OpenScore String Quartet for Optical Music Recognition (OSSQ-OMR), the first dataset dedicated to multi-part OMR. Built on the OpenScore String Quartet corpus, OSSQ-OMR pairs digitally encoded scores with their original scanned editions from IMSLP, with all images visually aligned to their transcriptions.\nThe dataset is released with score images at system and staff levels, and paired transcriptions in three encoding formats: Extended Linearized MusicXML (LMXE), **kern, and ABC. In total, OSSQ-OMR contains 24,544 system images and 98,172 staff images drawn from 116 string quartet scores. We accompany the dataset with a benchmark protocol and baseline results from two representative OMR models, evaluated across four random score-level splits with mutually exclusive test sets. Baselines reach OMR-NED as low as 3.6% on synthetic and 5.9% on scanned inputs; results reveal substantial effects of encoding and segmentation choices, with the LSTM-based baseline degrading on scanned inputs roughly 2.6× less than the Transformer-based baseline.","primarySubject":"Analysis tasks -> optical music recognition","secondarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases;MIR data fundamentals -> symbolic music processing","presentation":"Virtual","slackChannel":"p13-3-a-dataset-and","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7T9T6H","funFact":"","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_60.html","displayId":"P13-03","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Optical music recognition","Evaluation, datasets, and reproducibility","Novel datasets and use cases","MIR data fundamentals","Symbolic music processing"]},"83":{"uid":"83","title":"Low-Latency Real-Time Automatic Music Transcription with Band-Split RNN and Temporal Lookahead","authors":"Yuta Kusaka; Akira Maezawa","abstract":"This paper presents a real-time automatic music transcription (AMT) architecture that enables low-latency inference without compromising accuracy. Existing online AMT architectures exhibit a latency–accuracy trade-off because their convolutional encoders rely on lookahead to maintain transcription performance. To address this trade-off, we propose Band-Split RNN AMT, a low-latency architecture that processes sub-band features using a fully causal encoder based on recurrent neural networks. This RNN-based encoder requires no lookahead and effectively captures harmonic structure by modeling intra-band temporal and inter-band spectral dependencies. When additional latency is acceptable, the encoder can leverage future context through a flexible lookahead mechanism. In addition, we incorporate several architectural refinements and normalization techniques tailored to online AMT. Experiments on piano and guitar datasets show that the proposed model achieves a superior latency–accuracy trade-off relative to convolutional baselines. Under zero-lookahead configurations, our model improves the note-with-offset F1-score by 8.5 points while reducing latency by 34ms compared to the baselines, supporting practical real-time AMT even with latency below 100ms.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"","presentation":"Virtual","slackChannel":"p13-4-low-latency-real","channelUrl":"https://ismir2026.slack.com/archives/C0C6STGGM9N","funFact":"Yuta Kusaka is dedicated to training—both AI models and his own body! His goal is to bench press 100 kg.\n\nThe pianist in our demo is our colleague Rei Nishiyama, who also performed at last year’s ISMIR jam session.","flags":[],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_83.html","displayId":"P13-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Open Review"]},"452":{"uid":"452","title":"GlobalTap: Globally Diverse Beat Tracking Benchmark via Crowdsourced Tapping","authors":"William T. Botticelli-Wells; Manuel Anglada-Tort; Elif Celen; Rena Far; Harin Lee; Peter M. C. Harrison; Nori Jacoby","abstract":"Existing beat-tracking benchmarks draw from a narrow range of musical cultures, predominantly Western, and are difficult to scale due to reliance on expert annotation. We address both limits with GlobalTap, a benchmark of 200 popular-music excerpts spanning 59 countries, annotated by 746 participants recruited across ten countries. The benchmark is created using a scalable online tapping pipeline that records taps via audio for high temporal precision, aggregates pooled tap times into consensus beat annotations, and computes intrinsic confidence measures that flag perceptually difficult excerpts without requiring a reference. In a separate study on a different 240-excerpt validation set, tapped by an additional 752 participants, an independent group of 130 listeners rated how well the resulting beat tracks lined up with the music, and we detected no overall difference between our annotations and existing reference annotations. Crowd-derived annotations occasionally diverged from the references in phase and metrical interpretation, yet listeners did not consistently prefer the reference versions, suggesting that crowd-derived and existing reference annotations can reflect different perceptual or structural interpretations. The disagreements were often systematic rather than random, and metrical-level choices differed across the ten recruitment countries. We release GlobalTap together with open-source pipeline code as a listener-derived resource for evaluating beat-tracking systems on culturally diverse music.","primarySubject":"Musical features and properties -> rhythm, beat, tempo","secondarySubject":"Cognitive and user-centered MIR -> perception and cognition;Evaluation, datasets, and reproducibility -> annotation protocols;Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> novel datasets and use cases","presentation":"In-person","slackChannel":"p13-5-globaltap-globally-diverse","channelUrl":"https://ismir2026.slack.com/archives/C0C627X45RA","funFact":"William is a DJ and producer who recently moved to Ithaca from Amsterdam.\n\nIthaca gets very cold in the winter and has many waterfalls. Sometimes, the waterfalls even freeze.\n\nWilliam heard so much music from around the world while making GlobalTap that he began to believe he could understand many foreign languages. He couldn't.","flags":[],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_452.html","displayId":"P13-05","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Rhythm, beat, tempo","Cognitive and user-centered MIR","Perception and cognition","Evaluation, datasets, and reproducibility","Annotation protocols","Evaluation methodology","Novel datasets and use cases","Open Review"]},"258":{"uid":"258","title":"SymphonyGen: 3D Hierarchical Orchestral Generation with Controllable Harmony Skeleton","authors":"Xuzheng He; Nan Nan; Zhilin Wang; Ziyue Kang; Zhuoru Mo; Ao Li; Yu Pan; Xiaobing Li; Feng Yu; Xiaohong Guan","abstract":"Generating symphonic music requires simultaneously managing high-level structural form and dense, multi-track orchestration, yet existing symbolic models often struggle with a \"complexity-control imbalance\" between scalability and steerability. We present SymphonyGen, a 3D hierarchical framework for contemporary orchestral generation, whose cascading decoders decompose the bar, track, and event axes, keeping decoding memory far below flat token streams and enabling conditioning at every structural level. A beat-quantized multi-pitch harmony skeleton, which may be user-written, analyzed, or model-generated, provides \"short-score\" conditioning, enabling outline control while producing orchestral textures. The model is refined with reinforcement learning against a cross-modal acoustic reward from CLaMP 3 audio embeddings, and a dissonance-averse sampling algorithm suppresses unintended tonal clashes during inference. Objective evaluations show that both post-training mechanisms reduce dissonance while maintaining independent melodic metrics, and in subjective tests SymphonyGen is rated above baseline systems in quality and preference, significantly so among general listeners.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Applications -> music composition, performance, and production;Generative Tasks -> artistically-inspired generative tasks;Musical features and properties -> harmony, chords and tonality","presentation":"In-person","slackChannel":"p13-6-symphonygen-3d-hierarchical","channelUrl":"https://ismir2026.slack.com/archives/C0C60F3T1RP","funFact":"","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_258.html","displayId":"P13-06","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Applications","Music composition, performance, and production","Artistically-inspired generative tasks","Musical features and properties","Harmony, chords and tonality","Open Review"]},"269":{"uid":"269","title":"MPEcho: A Melody and Phoneme-Aware Generative Framework for Controllable Cover Song Generation","authors":"Wei-Jaw Lee; Hsuan-Yu Yeh; Ting-Yi Hu; Chih-Pin Tan; Fang-Duo Tsai; Yi-Hsuan Yang","abstract":"Cover song generation (CSG) should preserve the melodic and linguistic content of a reference song while recreating the remaining musical components. The state-of-the-art model SongEcho utilizes F0 sequences and voiced/unvoiced (V/UV) tags for conditioning; however, implicit linguistic information from V/UV tags cannot guarantee lyric accuracy, leading to a high phoneme error rate (PER). Inspired by singing voice synthesis (SVS), we propose MPEcho, which integrates a phoneme encoder and a length regulator (LR) into the SongEcho framework. By providing explicit phoneme-level conditioning and precise temporal boundaries, MPEcho significantly reduces PER. To enable this, we developed Phonsa, a Whisper-based automatic transcription model that provides high-precision phoneme-level annotations for singing voices, overcoming the scarcity of high-quality audio-phoneme pairs. Experimental results validate the effectiveness of Phonsa for alignment and MPEcho for end-to-end CSG. The audio samples, code and weights can be accessed from https://lonian6.github.io/MPEcho.github.io/.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Analysis tasks -> alignment, synchronization, and score following;MIR data fundamentals -> lyrics and other textual data;Musical features and properties -> melody and motives;Musical features and properties -> timbre, instrumentation, and singing voice","presentation":"","slackChannel":"p13-7-mpecho-a-melody","channelUrl":"https://ismir2026.slack.com/archives/C0C5T7Y84S1","funFact":"","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_269.html","displayId":"P13-07","presenterMode":"","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Analysis tasks","Alignment, synchronization, and score following","MIR data fundamentals","Lyrics and other textual data","Musical features and properties","Melody and motives","Timbre, instrumentation, and singing voice","Open Review"]},"298":{"uid":"298","title":"SKY-Piano: A Multimodal Piano Performance Dataset","authors":"Joonhyung Bae; Dawon Park; Taegyun Kwon; Yoon-Seok Choi; Hyeon Hur; Satoshi Obata; Shigeru Kai; Yohei Wada; Yu Takahashi; Akira Maezawa; Jaebum Park; Jonghwa Park; Juhan Nam","abstract":"Music information retrieval research on piano performance increasingly involves diverse modalities of data and annotations beyond audio and MIDI. We present SKY-Piano, a multimodal piano performance dataset that includes 11 hours of performance recordings of motion, multi-view video, audio, MIDI from 7 professional and 12 amateur pianists along with MusicXML scores. The performance pieces were selected considering playing technique, difficulty, and performer expertise on a shared core repertoire. The motion data include both hand and body motion, released in both flagged form, where samples lost to marker occlusion are marked as unreliable, and imputed form, where those gaps are reconstructed, together with Visual3D body-segment kinematics and other time-synchronized modalities. To easily browse different modalities of data at a glance, we provide an interactive web browser. In addition, we developed a fingering annotation model and tool for deriving pseudo fingering annotations from the MIDI and motion data. Lastly, we present MIDI-to-motion generation through a fine-tuning experiment as a use case of the dataset.","primarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","secondarySubject":"Evaluation, datasets, and reproducibility -> annotation protocols;MIR data fundamentals -> multimodality","presentation":"","slackChannel":"p13-8-sky-piano-a","channelUrl":"https://ismir2026.slack.com/archives/C0C60F51U2D","funFact":"","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_298.html","displayId":"P13-08","presenterMode":"","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Novel datasets and use cases","Annotation protocols","MIR data fundamentals","Multimodality","Open Review"]},"433":{"uid":"433","title":"Beyond Piano: Cross-Instrument MIDI Velocity Estimation via Differentiable SoundFont Proxies","authors":"Zhanhong He; Hanyu Meng; David Huang; Roberto Togneri","abstract":"Many music datasets contain MIDI notes but lack reliable velocities, defaulting to a constant value. This absence is especially problematic outside the piano domain, as velocity is a core component for expressive rendering, music generation, and performance analysis. This paper studies cross-instrument MIDI velocity estimation in this label-scarce setting. Starting from a piano-trained velocity estimator, we recast target-instrument adaptation as predicting renderer-conditioned velocities whose rendering matches the dynamics of the performance audio. This adaptation can be driven by either differentiable synthesizers (Diff-Synth) or our proposed differentiable SoundFont proxies (Diff-SFProxy). We highlight the Diff-SFProxy: it supervises velocity through note-wise, loudness-related acoustic parameters rather than waveform reconstruction, focusing gradients on velocity-dependent behavior. Experiments on piano and guitar show that Diff-SFProxy is effective for cross-instrument MIDI velocity estimation, while waveform-domain Diff-Synth degrades performance.","primarySubject":"Analysis tasks -> music transcription and annotation","secondarySubject":"Musical features and properties -> expression and performative aspects of music","presentation":"In-person","slackChannel":"p13-9-beyond-piano-cross","channelUrl":"https://ismir2026.slack.com/archives/C0C5T8GHUBF","funFact":"","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_433.html","displayId":"P13-09","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Analysis tasks","Music transcription and annotation","Musical features and properties","Expression and performative aspects of music"]},"414":{"uid":"414","title":"An Open Corpus for Arabic Maqam Recognition","authors":"Peter Traver; Layth Sidiq; Robert Rowe; Andrew J. Eisenberg","abstract":"Automatic recognition of maqam, the modal framework underlying much of Arabic music, remains a challenging problem for computational music analysis. While prior work on Turkish makam recognition has demonstrated promising results using pitch-class distribution (PCD) features and tonic-aware classifiers, these methods do not fully capture the nuances of performances in the Arabic maqam repertoire. This paper presents the Open Corpus for Arabic Maqam Recognition (OCMR) dataset, a curated collection of 1525 annotated recordings spanning 33 maqamat. We evaluate feature representations and classifiers for the maqam recognition task, adapting the Turkish-Ottoman Makam Music Analysis Toolbox (tomato) library from the UPF Barcelona CompMusic project for Arabic maqam. Re-implementation of tomato reproduces its original makam recognition accuracy on the OTMM benchmark. The adapted toolbox for Arabic maqam achieved an accuracy of 66.20\\% on the OCMR data. The low absolute percentages across all tasks suggest future work is needed to better understand maqam analysis. A panel discussion of Turkish \\& Arabic makam/maqam specialists was conducted and an outline of future technical research is proposed.","primarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","secondarySubject":"Applications -> music heritage and preservation;Computational musicology -> computational music theory and musicology;Evaluation, datasets, and reproducibility -> annotation protocols;Musical features and properties -> melody and motives;Musical features and properties -> representations of music","presentation":"","slackChannel":"p13-10-an-open-corpus","channelUrl":"https://ismir2026.slack.com/archives/C0C627TAETW","funFact":"","flags":["student"],"session":"PS-13","posterUrl":"https://ismir2026program.ismir.net/poster_414.html","displayId":"P13-10","presenterMode":"","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Novel datasets and use cases","Applications","Music heritage and preservation","Computational musicology","Computational music theory and musicology","Annotation protocols","Musical features and properties","Melody and motives","Representations of music","Open Review"]},"125":{"uid":"125","title":"Estimating the Reliability of Dynamic Time Warping Alignments Using Circumstantial Evidence","authors":"Aanya Pratapneni; Alice Yuan; TJ Tsai","abstract":"Recent works have explored ways to handle uncertainty in dynamic time warping (DTW) alignment paths through the use of differentiable variants of DTW like Soft-DTW. In this paper, we approach the issue of uncertainty in DTW alignment paths in a different way. Given a DTW alignment path, we propose a metric that indicates how reliable a local segment of the alignment path is. The intuition for our metric is based on the idea of circumstantial evidence. If DTW has found a very prominent path, then if we re-run the alignment with relaxed boundary conditions, it will still pick the same path. If, on the other hand, DTW has found a “weak” path, then re-running the alignment with relaxed boundary conditions will likely yield a different path. Accordingly, our reliability metric is computed by picking a local section of the DTW alignment path, re-estimating the alignment with FlexDTW (which allows flexibility in the boundary conditions), and then measuring how well the DTW and FlexDTW paths agree. We assess the proposed reliability metric on DTW alignment paths containing both matching and non-matching regions across a range of scenarios on an audio-audio alignment task. We find that the reliability metric correctly identifies reliable regions of the alignment path with an aggregate AUROC of 0.97. This approach provides an unsupervised method for estimating the reliability of a DTW alignment path.","primarySubject":"Analysis tasks -> alignment, synchronization, and score following","secondarySubject":"","presentation":"Virtual","slackChannel":"p14-1-estimating-the-reliability","channelUrl":"https://ismir2026.slack.com/archives/C0C5H5UMXM5","funFact":"","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_125.html","displayId":"P14-01","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Analysis tasks","Alignment, synchronization, and score following","Open Review"]},"525":{"uid":"525","title":"Music Restoration via Latent Operator Optimization and Diffusion Model Priors","authors":"Michal Švento; Eloi Moliner; Valtteri Kallinen; Lauri Juvela; Vesa Välimäki; Pavel Rajmic","abstract":"Music restoration seeks to recover a clean signal from an observed recording degraded by an unknown effect, distortion, or corruption. Existing systems often rely on paired training data and distortion-specific supervision, which limits their use when the forward process is not known in advance. We propose LOUDAR (Latent-space Optimization of Unknown Distortion for Audio Restoration) a general-purpose restoration method that operates in the latent space of a pretrained audio autoencoder and models the unknown distortion as a learnable latent operator. At inference time, LOUDAR alternates between estimating the clean latent variable and updating the latent operator parameters. An unconditional latent diffusion model provides a prior over clean audio and regularizes this inference by steering the latent estimate toward the manifold of clean recordings. Because the degradation model is adapted per input, the approach is broadly applicable across diverse restoration problems.\nWe evaluate LOUDAR on singing voice effect removal and restoration, as well as guitar distortion removal, and show that it consistently improves over degraded inputs and is competitive with supervised and unsupervised baselines in waveform and latent domains.","primarySubject":"Generative Tasks -> transformations","secondarySubject":"MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p14-2-music-restoration-via","channelUrl":"https://ismir2026.slack.com/archives/C0C5SB3H1NF","funFact":"","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_525.html","displayId":"P14-02","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Generative Tasks","Transformations","MIR data fundamentals","Music signal processing","Open Review"]},"527":{"uid":"527","title":"Music Inpainting with Similarity-Guided Latent Diffusion","authors":"Sean Turland; Eloi Moliner; Vesa Välimäki","abstract":"A music inpainting technique combining similarity matching and a diffusion-based generative model is proposed to replace large segments of audio recordings with suitable alternatives. The proposed hybrid method, Similarity-Guided Diffusion Posterior Sampling (SimDPS), retrieves candidate replacement segments from the wider signal based on contextual similarity and uses them to guide the diffusion process toward coherent reconstructions. Being model-agnostic, the framework is compatible with any pre-trained diffusion model operating in either the waveform or latent domain. Objective evaluations demonstrate that SimDPS outperforms widely adopted RePaint and DPS methods across acoustic features and neural-embedding distance metrics. Subjective evaluations on multi-genre music datasets confirm that SimDPS enhances perceptual plausibility compared to previous diffusion-based methods and demonstrates robustness to poorly matched similarity candidates. These findings pave the way for improved music restoration and context-aware audio editing.","primarySubject":"Generative Tasks -> music and audio synthesis","secondarySubject":"Analysis tasks -> similarity metrics;MIR data fundamentals -> music signal processing","presentation":"In-person","slackChannel":"p14-3-music-inpainting-with","channelUrl":"https://ismir2026.slack.com/archives/C0C5SB2N8P5","funFact":"Aspects of this work were inspired by ideas used in Richard D. James' \"SampleBrain\" software. \nSean Turland enjoys various forms of experimental music, and would love to chat with any researchers applying their technology to their art.","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_527.html","displayId":"P14-03","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Music and audio synthesis","Analysis tasks","Similarity metrics","MIR data fundamentals","Music signal processing","Open Review"]},"539":{"uid":"539","title":"Assessing AI-Generated Music Detection in Real-World Broadcast Monitoring","authors":"David López-Ayala; Fernando Garcia de la Cruz; Pablo Zinemanas; Emilio Molina; Martín Rocamora","abstract":"The proliferation of AI-generated music in broadcast media raises concerns about transparency and fair compensation, but reliable detection under real broadcast conditions remains unresolved. Existing studies report substantial performance degradation in this domain, yet their evaluations are limited to synthetic broadcast data. To address this gap, we introduce BAMM (Broadcast AI-Music Monitoring), a 40-hour dataset of real-world television recordings containing AI-generated and human-made music. We compare clean-trained and broadcast-trained CNN variants across three progressively more challenging scenarios: Clean Foreground Music (CFM), Synthetic TV Broadcast (STB), and Real TV Broadcast (RTB). Both models achieve near-perfect performance on CFM but degrade substantially under synthetic broadcast conditions. Broadcast-oriented training improves robustness compared with clean training, although performance remains limited. On RTB, evaluated using BAMM, both models degrade further and show substantial score overlap between AI-generated and human-made music. These results expose a critical domain gap and show that current training approaches on CNN-based detectors remain insufficient for reliable AI-generated music detection in broadcast monitoring.","primarySubject":"Evaluation, datasets, and reproducibility -> novel datasets and use cases","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology;Evaluation, datasets, and reproducibility -> evaluation metrics;Generative Tasks -> music and audio synthesis;MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p14-4-assessing-ai-generated","channelUrl":"https://ismir2026.slack.com/archives/C0C5SB3KF5Z","funFact":"1) BMAT is a spinoff of UPF-MTG.\n2) Half of the authors are Mexican, half are Uruguayan, and one is Spanish.","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_539.html","displayId":"P14-04","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Evaluation, datasets, and reproducibility","Novel datasets and use cases","Evaluation methodology","Evaluation metrics","Generative Tasks","Music and audio synthesis","MIR data fundamentals","Music signal processing"]},"552":{"uid":"552","title":"InvFlowFD: Reference-Free and Background-Set-Free Perceptual Music Quality Metric with Flow Matching Inversion","authors":"Alon Ziv; Harel Pogoda; Yossi Adi","abstract":"Existing reference-free methods for evaluating music perceptual quality alleviate the need for paired noisy-clean data, but they still rely on a background set, which is used to compute aggregated statistics of clean audio samples. In this work, we propose a novel approach that eliminates this requirement, achieving background-set-free and reference-free quality estimation using only a pre-trained Flow Matching backbone. We demonstrate that unconditional Flow Matching inversion via simple Euler integration is sufficient to detect various artificial distortions and accurately rank music generation models against human perceptual judgments. We introduce InvFlowFD, which performs flow inversion and compares a group of inverted samples to the prior distribution. We evaluate our method against prior work, quantitatively and with a thorough human study. Results suggest that InvFlowFD is highly correlated with human perception of sound distortions, as well as generative models' quality, while being more flexible and less restrictive than existing metrics.","primarySubject":"Generative Tasks -> evaluation metrics","secondarySubject":"Generative Tasks -> qualitative evaluations","presentation":"In-person","slackChannel":"p14-5-invflowfd-reference-free","channelUrl":"https://ismir2026.slack.com/archives/C0C62868BL4","funFact":"Apart from CS research, Harel Pogoda produces House music and has over 3M Spotify streams:)","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_552.html","displayId":"P14-05","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Generative Tasks","Evaluation metrics","Qualitative evaluations","Open Review"]},"335":{"uid":"335","title":"Culture-Specific Computational Musicology in Practice: DiArMaqAr (Digital Arabic Maqām Archive)","authors":"Khyam Allami; Ibrahim El Khansa; Mohammad El Asal; Joseph Bakarji","abstract":"Culture-specific computational musicology approaches a musical culture as a system of interconnected entities requiring adequate digital traces and rigorous formalisation grounded in the culture’s own theoretical frameworks. For Arabic maqām and related modal-music traditions, the gap between documented theory and computationally tractable infrastructure identified by Gedik and Bozkurt has persisted for decades. This paper presents the first practical implementation of culture-specific computational musicology for Arabic maqām, the Digital Arabic Maqām Archive (DiArMaqAr).\n\nThe multilingual archive brings together historical tanāghīm (tuning systems), ajnās (tri/tetra/pentachords), maqāmāt (modal frameworks), suyūr (melodic pathways), and intiqālāt (modulations) in an open-access interactive web application and corpus, with full bibliographic attribution. All entities are stored as sequences of tuning-system-independent Persian-Arab-Ottoman (PAO) note names, making transposition and rendering across tuning systems computable operations. Aleppine violinist Sāmī al-Shawwā’s 1946 intiqālāt guidelines, the only rule-based treatment of maqām modulation in written theory, are formalised algorithmically for the first time and validated against the modulation targets given by his contemporary Tawfīq al-Ṣabbāgh. \n\nAlongside the archive’s design and theoretical grounding, the paper introduces Tanghīm, a downstream open-source cross-platform MIDI plugin that draws on the platform’s REST API to bring accurately tuned maqāmāt into digital audio workstations for music creatives. The paper closes with an analysis of the costs, yields, and portability of the approach to modal-music traditions where the Gedik-Bozkurt gap remains open.","primarySubject":"Computational musicology","secondarySubject":"Applications -> digital libraries and archives;Creativity -> tools for artists;Evaluation, datasets, and reproducibility -> novel datasets and use cases;Musical features and properties -> representations of music;Responsibility and Ethics in MIR -> philosophical and methodological foundations","presentation":"In-person","slackChannel":"p14-6-culture-specific-computational","channelUrl":"https://ismir2026.slack.com/archives/C0C60F98X7T","funFact":"","flags":[],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_335.html","displayId":"P14-06","presenterMode":"onsite","materials":["paper"],"keywords":["Computational musicology","Applications","Digital libraries and archives","Creativity","Tools for artists","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Musical features and properties","Representations of music","Responsibility and Ethics in MIR","Philosophical and methodological foundations"]},"472":{"uid":"472","title":"Musicologically Interpretable Deep Learning Framework for Arabic Maqam Identification and Analysis","authors":"Joseph Bakarji; Khyam Allami; Mohammad El Asal; Dany Abou Jaoude","abstract":"Arabic maqams are melodic modes built from characteristic intervals that vary across regional traditions and do not conform to equal temperament or a fixed tuning reference frequency, making their computational analysis and identification from audio challenging. We present a musicologically informed deep learning framework that integrates Arabic maqam theory into both the preprocessing pipeline and the interpretation of learned representations. Using maqam, tuning, and interval templates from a Digital Arabic Maqam Archive (Allami et al, 2026), we first show that tonic estimation is the dominant bottleneck in maqam identification where standard classifiers approach an accuracy ceiling. We thus introduce MaqamNet, a circular-convolution residual network that bypasses explicit tonic estimation by encoding the tonic as a latent shift variable in the cross-correlation between pitch histograms and theory-derived templates. Evaluated on a unified corpus of 619 recordings from three Arabic datasets (Oud taqasim, Quranic recitation (Shahriar et al., 2021), and Cairo Congress 1932 (Bozkurt, 2025)), MaqamNet achieves high accuracy from pitch alone. Ablations show that the circular-convolution architecture, rather than the theory-derived template initialization, drives this performance. However, the templates provide musicological interpretability, with Grad-CAM revealing theory-aligned pitch and temporal patterns.","primarySubject":"Musical features and properties -> representation learning","secondarySubject":"Analysis tasks -> automatic classification;Computational musicology -> computational music theory and musicology","presentation":"In-person","slackChannel":"p14-7-musicologically-interpretable-deep","channelUrl":"https://ismir2026.slack.com/archives/C0C5SAU6TJ7","funFact":"","flags":[],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_472.html","displayId":"P14-07","presenterMode":"onsite","materials":["paper","poster","video"],"keywords":["Musical features and properties","Representation learning","Analysis tasks","Automatic classification","Computational musicology","Computational music theory and musicology","Open Review"]},"518":{"uid":"518","title":"How Much AI Is in This Track? Quantifying the Proportion of AI-Generated Stems in Hybrid Music Mixtures","authors":"Fernando Garcia de la Cruz; David López-Ayala; Pablo Zinemanas; Emilio Molina; Martín Rocamora","abstract":"AI-generated music is increasingly used at the stem level, with producers integrating synthetic drums, basslines, or vocals alongside human-performed instruments. However, current AI music detection systems are binary, treating tracks as either fully AI or fully human. In this paper, we reformulate AI music detection as a regression problem on a continuous AI energy ratio, alpha in [0, 1]. We propose a methodology that leverages a multi-track music dataset to assemble mixtures of human-performed and AI-reconstructed stems (obtained using a neural audio codec) with known proportions of each content type. Using this approach, we first show that a CNN-based model trained on fully AI-generated or human-performed tracks, which achieves >99% accuracy as a binary detector, when faced with mixed content, yields an output that rises with the AI stems' energy contribution, acting as a noisy and miscalibrated estimator. Our analysis of the influence of different stems shows that detection sensitivity depends on the instrument and reflects its frequency content: drums and guitar carry strong codec-artifact signatures, while vocals and bass are less detectable. Based on these insights, we train a similar CNN-based model for regression of alpha, achieving MAE = 0.076 and R^2 = 0.85 on held-out mixtures from the same pipeline. These results suggest that the regression formulation is an initial promising step towards AI-music detection in realistic music production workflows.","primarySubject":"MIR data fundamentals -> music signal processing","secondarySubject":"Evaluation, datasets, and reproducibility -> evaluation methodology;Generative Tasks -> evaluation metrics;Generative Tasks -> music and audio synthesis","presentation":"Virtual","slackChannel":"p14-8-how-much-ai","channelUrl":"https://ismir2026.slack.com/archives/C0C60FT00MP","funFact":"🎚️ A detector with more than 99% accuracy on fully-AI vs. fully-human tracks scores very bad when asked \"how much AI?\" on hybrid mixes.\n\n🤖 Deezer reports about 90,000 fully AI-generated uploads a day, over 50% of new uploads at the June 2026 peak, yet only 1-3% of streams. This paper tackles the tracks in between: the ones that are partly AI.\n\n🚀🐴 The cover astronaut nods to the classic \"astronaut riding a horse\" prompt that showed off DALL-E 2 in 2022. The irony: the cover art and thumbnail are partly AI-generated too, so \"How much AI is is this image?\".","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_518.html","displayId":"P14-08","presenterMode":"online","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Music signal processing","Evaluation, datasets, and reproducibility","Evaluation methodology","Generative Tasks","Evaluation metrics","Music and audio synthesis"]},"530":{"uid":"530","title":"MetaMatch: Reconciling Metadata for Aggregating Music Difficulty Annotations","authors":"Megan E. Finch; Chris G. Willcocks; Mark R. H. Gotham","abstract":"The inconsistency of metadata across sources is a persistent problem for large-scale music research. The variety of approaches to documenting even basic elements such as artists’ names and work titles makes it difficult to aggregate data across different sources. This paper presents a comparative analysis of approaches to this problem, combining the OpenRefine Reconciliation pipeline with LLM-based strategies for entity matching. We also explore one potential application of this pipeline in sheet music difficulty classification. By gathering difficulty labels from established music exam board syllabi and matching them to corresponding IMSLP entries, we link expert difficulty annotations to sheet music available through IMSLP for automatic sheet music difficulty classification. The resulting piano dataset consists of 7,737 repertoire-list entries, of which 2,389 are automatically matched to 1,210 distinct IMSLP work identifiers. Our method produces data that combine high-quality expert annotations with the scale more typically associated with crowdsourced corpora, and can be extended to additional instruments.\n\nWe release the matching pipeline, dataset, and code at https://meganelisabethfinch.github.io/metamatch to support further work on cross-corpus music analysis, both for this use case and for the many others which depend on robust metadata alignment at scale.","primarySubject":"MIR data fundamentals -> metadata, tags, linked data, and semantic web","secondarySubject":"Computational musicology -> computational music theory and musicology;Computational musicology -> digital musicology;Evaluation, datasets, and reproducibility -> novel datasets and use cases;MIR data fundamentals -> lyrics and other textual data","presentation":"","slackChannel":"p14-9-metamatch-reconciling-metadata","channelUrl":"https://ismir2026.slack.com/archives/C0C5WK00YUW","funFact":"The Computer Science department at Durham is frequented by crows (mostly because Chris feeds them...). Megan plays mandolin, baritone horn, and piano.","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_530.html","displayId":"P14-09","presenterMode":"","materials":["paper","poster","video"],"keywords":["MIR data fundamentals","Metadata, tags, linked data, and semantic web","Computational musicology","Computational music theory and musicology","Digital musicology","Evaluation, datasets, and reproducibility","Novel datasets and use cases","Lyrics and other textual data"]},"320":{"uid":"320","title":"Explore This! Beat and Downbeat Tracking From a Learned Tatum Grid","authors":"Robert Kihlborg; André Holzapfel; Jan Schlüter","abstract":"State-of-the-art beat tracking is based on neural networks that model beats and downbeats as individual events, often followed by a separate Dynamic Bayesian Network (DBN) enforcing musical consistency. We propose a novel architecture that combines deep learning with differentiable theory-based modules: Specifically, we design a system that first predicts a tatum grid, then groups tatum events to beats and downbeats, explicitly modeling multiple options for the current tempo and time signature in a bar-pointer like fashion. Our approach improves continuity metrics compared to existing DBN-free systems and is more versatile than the de-facto standard DBN, but still falls behind the best DBN-based methods on average. Given the many design choices and hyperparameters involved, we assume we have not yet explored the full potential of this idea and publish our code and invite others to explore this further.","primarySubject":"Musical features and properties -> rhythm, beat, tempo","secondarySubject":"MIR data fundamentals -> music signal processing","presentation":"Virtual","slackChannel":"p14-10-explore-this-beat","channelUrl":"https://ismir2026.slack.com/archives/C0C5H62T58F","funFact":"Robert Kihlborg likes music with a lot of time signature changes, which is still very difficult for even the best Beat Trackers. This helped spark an interest in improving performance on such pieces.","flags":["student"],"session":"PS-14","posterUrl":"https://ismir2026program.ismir.net/poster_320.html","displayId":"P14-10","presenterMode":"online","materials":["paper","poster","video"],"keywords":["Musical features and properties","Rhythm, beat, tempo","MIR data fundamentals","Music signal processing","Open Review"]}},"lbdSession":{"id":"LBD","slots":[{"kind":"lbd","date":"2026-11-12","start":"15:00","end":"17:00"}]},"lbds":[{"uid":"569","displayId":"LBD-569","title":"GroovyUI: A Patch-Bay for Local AI Audio Exploration","authors":"Yuan-Yi Fan","channelUrl":"https://ismir2026.slack.com/archives/C0CACMPGBHQ","pageUrl":"https://ismir2026program.ismir.net/lbd_569.html","presenterMode":"online","keywords":["Artisanal Intelligence","Creative Agents","Al Trust","Modular Synth","Human-AI Interaction"],"materials":["paper","poster","video"]},{"uid":"570","displayId":"LBD-570","title":"Integrated Multi-Effect Parameter Prediction and Guitar Effects Simulation","authors":"Aarush Gupta; Lior Arbel","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ5MQTE","pageUrl":"https://ismir2026program.ismir.net/lbd_570.html","presenterMode":"onsite","keywords":["Guitar Tone Modeling","Digital Signal Processing (DSP)","FAUST DSP Framework","JUCE framework","VST3/AU plugin"],"materials":["paper","poster","video"]},{"uid":"571","displayId":"LBD-571","title":"Where Does Genre Stop Following the Chords? A Fixed-Frame Listening Probe of Bounded Chord-Symbol Genre Identity (TheArtist)","authors":"Jinju Lee","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LELRPD","pageUrl":"https://ismir2026program.ismir.net/lbd_571.html","presenterMode":"online","keywords":["chord progressions","genre adaptation","symbolic music generation","LoRA fine-tuning","listening probe"],"materials":["paper","poster","video"]},{"uid":"575","displayId":"LBD-575","title":"IdioMap: Passage-Level Composer Evidence for Symbolic Piano Performance","authors":"Jaeseong Heo; Sangkyoo Park; Joonhyung Bae","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JM1329","pageUrl":"https://ismir2026program.ismir.net/lbd_575.html","presenterMode":"online","keywords":["composer identification","passage-level evidence","symbolic piano performance","musical influence","interpretable music classification"],"materials":["paper","poster","video"]},{"uid":"576","displayId":"LBD-576","title":"Active Silence in Beethoven: Low-Energy Pause Strategies in the Grave of the Pathétique","authors":"Harold Wang","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ5H18C","pageUrl":"https://ismir2026program.ismir.net/lbd_576.html","presenterMode":"online","keywords":["expressive performance analysis","musical pauses","piano performance","audio signal analysis"],"materials":["paper","poster","video"]},{"uid":"577","displayId":"LBD-577","title":"Navigable texture spaces for the analysis of rhythm guitar tablature corpora","authors":"Luka Desfarges; Louis Bigo; Léo Dupouey","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JLGS7P","pageUrl":"https://ismir2026program.ismir.net/lbd_577.html","presenterMode":"online","keywords":["Guitar","Tablature","Web","Plot","Corpus"],"materials":["paper","poster","video"]},{"uid":"578","displayId":"LBD-578","title":"BowDET: An Interactive Toolkit for Audio-Based Bow-Change Detection","authors":"Haotian Yuan; Yushi Su","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LF9NUB","pageUrl":"https://ismir2026program.ismir.net/lbd_578.html","presenterMode":"online","keywords":["bowed string performance","bow-change detection","music information retrieval","spectral boundary detection","interactive annotation toolkit"],"materials":["paper","poster","video"]},{"uid":"579","displayId":"LBD-579","title":"Soprano Lab: Measure-Aware Question Answering over Expert Vocal-Performance Annotations","authors":"Minhee Lee; Juhan Nam","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUE3ACE","pageUrl":"https://ismir2026program.ismir.net/lbd_579.html","presenterMode":"online","keywords":["Music education","Retrieval-augmented generation (RAG)","Question answering","Expert annotation"],"materials":["paper","poster","video"]},{"uid":"580","displayId":"LBD-580","title":"PANEL: An Open-Source, Self-Hosted Web Platform for Human Evaluation of Generative Models","authors":"Matteo Spanio; Andrea Poltronieri; Martín Rocamora","channelUrl":"https://ismir2026.slack.com/archives/C0CACMQT9R8","pageUrl":"https://ismir2026program.ismir.net/lbd_580.html","presenterMode":"online","keywords":["evaluation","listening tests","open-source software","reproducibility"],"materials":["paper","poster","video"]},{"uid":"583","displayId":"LBD-583","title":"trackswitch2: Interactive Playback Across Aligned Music Timelines","authors":"Manuel Peters; Stefan Balke; Nils Werner; Fabian-Robert Stöter; Bernd Edler; Meinard Müller","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ6PQRJ","pageUrl":"https://ismir2026program.ismir.net/lbd_583.html","presenterMode":"online","keywords":["interface","software","alignment","playback","trackswitch"],"materials":["paper","poster","video"]},{"uid":"585","displayId":"LBD-585","title":"Few-Shot Reconstruction of Microphone Directivity Fields with Cross-Microphone Priors","authors":"Haochen Wang; Juncheng Yu","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ5U48L","pageUrl":"https://ismir2026program.ismir.net/lbd_585.html","presenterMode":"online","keywords":["microphone directivity","few-shot reconstruction","cross-microphone priors","sampling geometry","data augmentation"],"materials":["paper","poster","video"]},{"uid":"586","displayId":"LBD-586","title":"Beyond Onset Detection: Unsupervised Note Boundary Detection for Bowed Strings","authors":"Haotian Yuan; Yushi Su","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ5RACC","pageUrl":"https://ismir2026program.ismir.net/lbd_586.html","presenterMode":"online","keywords":["Note boundary detection","bowed string instruments","unsupervised methods","spectral difference","onset detection"],"materials":["paper","poster","video"]},{"uid":"587","displayId":"LBD-587","title":"PawChorale: A Scalable Synthetic Dataset for Part-Aware Choral Transcription and Separation","authors":"Hanyu Meng; Zhanhong He; Kaifeng Li; Yaolong Ju","channelUrl":"https://ismir2026.slack.com/archives/C0C9XG4GX7Z","pageUrl":"https://ismir2026program.ismir.net/lbd_587.html","presenterMode":"onsite","keywords":["Synthetic choral dataset","Singing voice synthesis","Part-aware choral transcription","Choral source separation","Audio–MIDI alignment"],"materials":["paper","poster","video"]},{"uid":"588","displayId":"LBD-588","title":"Bring Your Own Music: Towards Community-Built, On-Demand Benchmarks of MLLM Music Perception","authors":"Tomáš Sourada; Katia Vendrame; Jan Hajič jr.","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ6PJCU","pageUrl":"https://ismir2026program.ismir.net/lbd_588.html","presenterMode":"online","keywords":["benchmark","LLM","multimodality","music perception"],"materials":["paper","poster","video"]},{"uid":"589","displayId":"LBD-589","title":"FP-JEPA: Toward Robust Music Fingerprinting With Joint-Embedding Prediction","authors":"Harishkumar Kishorkumar Prajapati; Mark Sandler","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUEQWTY","pageUrl":"https://ismir2026program.ismir.net/lbd_589.html","presenterMode":"online","keywords":["audio fingerprinting","music identification","robustness","self-supervised learning","joint-embedding predictive architecture (JEPA)","foundation models","metric learning"],"materials":["paper","poster","video"]},{"uid":"590","displayId":"LBD-590","title":"Reflector: Arrangement-Aware Harmonic Retrieval for Sample-Based Composition","authors":"Austin Rockman","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ6P48L","pageUrl":"https://ismir2026program.ismir.net/lbd_590.html","presenterMode":"online","keywords":["sample retrieval","harmonic compatibility","contrastive representation learning","music creation tools"],"materials":["paper","poster","video"]},{"uid":"591","displayId":"LBD-591","title":"MusExam: A Blueprint-Driven Architecture for Reproducible Music-Theory Learning and Mock Exams","authors":"Hale Fulya Çelikel Soğancı","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LG95DH","pageUrl":"https://ismir2026program.ismir.net/lbd_591.html","presenterMode":"onsite","keywords":["Music Theory","Browser Native MIR","International Graded Theory Exams","Parametric Exam Generation"],"materials":["paper","poster","video"]},{"uid":"596","displayId":"LBD-596","title":"Building a Dataset for Music Sample Identification","authors":"R. Oguz Araz; Xavier Lizarraga; Xavier Serra; Dmitry Bogdanov","channelUrl":"https://ismir2026.slack.com/archives/C0CB78RJDK2","pageUrl":"https://ismir2026program.ismir.net/lbd_596.html","presenterMode":"online","keywords":["music sample identification","dataset","graph-based data splitting","music information retrieval"],"materials":["paper","poster","video"]},{"uid":"597","displayId":"LBD-597","title":"Vector Arithmetic on Audio Effects in Contrastive Embeddings Space","authors":"Paula Lauren","channelUrl":"https://ismir2026.slack.com/archives/C0CB78SM68G","pageUrl":"https://ismir2026program.ismir.net/lbd_597.html","presenterMode":"online","keywords":["audio embeddings","CLAP","audio effects","vector arithmetic"],"materials":["paper","poster","video"]},{"uid":"599","displayId":"LBD-599","title":"Controlled Stochastic Variation of a Galician Folk Melody: A Relative-Displacement Demo","authors":"Xairo Campos Blanco","channelUrl":"https://ismir2026.slack.com/archives/C0CB78RQHEU","pageUrl":"https://ismir2026program.ismir.net/lbd_599.html","presenterMode":"online","keywords":["Algorithmic composition","Markov chains","Symbolic music generation","Galician folk music","Stochastic music"],"materials":["paper","poster","video"]},{"uid":"601","displayId":"LBD-601","title":"Suite José Suárez: Audience-Directed Visual Attention for Networked Open Scores","authors":"Xairo Campos Blanco","channelUrl":"https://ismir2026.slack.com/archives/C0CB78SVBB2","pageUrl":"https://ismir2026program.ismir.net/lbd_601.html","presenterMode":"online","keywords":["Audience participation","Interactive music systems","Image-to-music mapping","Open scores","Networked performance"],"materials":["paper","poster","video"]},{"uid":"602","displayId":"LBD-602","title":"HarmonyForge: A Deterministic Co-Creative System for Symbolic Music Arrangement","authors":"Dulf Vincent Genis; Shivam Patel; Muawiz Khan; Yiren Liu; Jeff Ginger; Yun Huang","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LHB491","pageUrl":"https://ismir2026program.ismir.net/lbd_602.html","presenterMode":"onsite","keywords":["Symbolic Music Arrangement","Deterministic Constraint Optimization","Explainable AI","Human-AI Co-Creation","Voice-Leading Constraints"],"materials":["paper","poster","video"]},{"uid":"603","displayId":"LBD-603","title":"Do Game-Theoretic Decision Rules Improve Two-Voice Music Generation?","authors":"Po-Chen Liu; Hsiang-Yen Fan","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUFBNES","pageUrl":"https://ismir2026program.ismir.net/lbd_603.html","presenterMode":"online","keywords":["Symbolic Music Generation","Game Theory","Polyphonic Music Generation","Game-Theoretic Decision Rules","Perceptual Evaluation"],"materials":["paper","poster","video"]},{"uid":"604","displayId":"LBD-604","title":"Source-Constrained Neural-Assisted Resampling for Concatenative Singing Synthesis","authors":"Yang Yang","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUG8VN2","pageUrl":"https://ismir2026program.ismir.net/lbd_604.html","presenterMode":"online","keywords":["singing voice synthesis","concatenative synthesis","source fidelity","neural-assisted resampling","DDSP"],"materials":["paper","poster","video"]},{"uid":"605","displayId":"LBD-605","title":"KarMIDI: A Symbolic Dataset for Carnatic Melody Harmonization","authors":"Nagamanoj Karunakaran; Ajay Srinivasamurthy; Arti Arya","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JPCSMB","pageUrl":"https://ismir2026program.ismir.net/lbd_605.html","presenterMode":"online","keywords":["Carnatic Music","Symbolic Music Dataset","Melody Harmonization","Music Information Retrieval (MIR)","BiLSTM"],"materials":["paper","poster","video"]},{"uid":"608","displayId":"LBD-608","title":"ScorePrompts: Natural-Language Exploration of Symbolic Music Scores through Analysis","authors":"Emmanouil Karystinaios; Gerhard Widmer","channelUrl":"https://ismir2026.slack.com/archives/C0CACMS8YFL","pageUrl":"https://ismir2026program.ismir.net/lbd_608.html","presenterMode":"online","keywords":["Score Prompts","LLM orchestration","Music analysis","grounded descriptions","information retrieval"],"materials":["paper","poster","video"]},{"uid":"609","displayId":"LBD-609","title":"Pocket Producer: Session-Conditioned Retrieval as User-Controlled Proposal","authors":"Chuling Li","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LHB2UB","pageUrl":"https://ismir2026program.ismir.net/lbd_609.html","presenterMode":"online","keywords":["music production","content-based retrieval","weak supervision","ranking evaluation","human-computer interaction"],"materials":["paper","poster","video"]},{"uid":"610","displayId":"LBD-610","title":"Measuring the Music Itself","authors":"Colin Munroe","channelUrl":"https://ismir2026.slack.com/archives/C0CB78T22LQ","pageUrl":"https://ismir2026program.ismir.net/lbd_610.html","presenterMode":"online","keywords":["Human annotation","Music datasets","Popular music analysis","Longitudinal music analysis","Music production"],"materials":["paper","poster","video"]},{"uid":"611","displayId":"LBD-611","title":"SpliceGen: Multi-Track Loop Generation Using Flow Matching And Inversion With Non-Parallel Data","authors":"Alejandro Koretzky; Amantur Amatov; Anastasiia Herus; Andy Thompson; Daniel Gonzalez; Jeff Sontag; Julian Lenz; Kyle Billings; Nate Raw; Naveen Sasalu Rajashekharappa; Shahan Nercessian; Yurii Bondarenko","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LHEH99","pageUrl":"https://ismir2026program.ismir.net/lbd_611.html","presenterMode":"online","keywords":["neural audio resynthesis","musical audio generation","stem generation","flow matching","RF inversion"],"materials":["paper","poster","video"]},{"uid":"612","displayId":"LBD-612","title":"A User-Centered View of AI Music Description","authors":"Eun Ji Oh; Sangheon Park; Yonghyun Kim; Jun Woo Beck","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LHG7DZ","pageUrl":"https://ismir2026program.ismir.net/lbd_612.html","presenterMode":"online","keywords":["music description","music captioning","user-centered MIR","survey research"],"materials":["paper","poster","video"]},{"uid":"613","displayId":"LBD-613","title":"Quill: Learning Timbre and Performance Behavior as Separate, Swappable Assets","authors":"Chuling Li","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ894SG","pageUrl":"https://ismir2026program.ismir.net/lbd_613.html","presenterMode":"online","keywords":["music and audio synthesis","timbre","instrumentation","and singing voice","expression and performative aspects of music","evaluation methodology","representations of music"],"materials":["paper","poster","video"]},{"uid":"615","displayId":"LBD-615","title":"TSM Diff: A new parallel-tree-traversal algorithm for the comparison of music score files","authors":"Xiaobo Wang; Florent Jacquemard","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ8SE2G","pageUrl":"https://ismir2026program.ismir.net/lbd_615.html","presenterMode":"online","keywords":["symbolic music processing","representation of music notation","digital music score diffing","similarity metrics for music notation","digital libraries and archives"],"materials":["paper","poster","video"]},{"uid":"616","displayId":"LBD-616","title":"Probing Pretrained Symbolic and Audio Models for Perceptual Qualities of Piano Performance","authors":"Eli Mason","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQASY1F","pageUrl":"https://ismir2026program.ismir.net/lbd_616.html","presenterMode":"online","keywords":["linear probing","evaluation","performance analysis"],"materials":["paper","poster","video"]},{"uid":"617","displayId":"LBD-617","title":"Broad-Band F0 Modulation Emotion Transfer from Singing to Speech","authors":"Mounia Drissi","channelUrl":"https://ismir2026.slack.com/archives/C0C9XG6N14P","pageUrl":"https://ismir2026program.ismir.net/lbd_617.html","presenterMode":"online","keywords":["Cross-domain emotion transfer","Singing voice analysis","Music emotion recognition","Voice production features"],"materials":["paper","poster","video"]},{"uid":"618","displayId":"LBD-618","title":"Robust Pitch, Alignment, and Technique Recognition for Interpretable Erhu Performance Assessment","authors":"Yongchang Gan","channelUrl":"https://ismir2026.slack.com/archives/C0CB78TG68G","pageUrl":"https://ismir2026program.ismir.net/lbd_618.html","presenterMode":"online","keywords":["Erhu","music performance analysis","score-event alignment","pitch tracking","technique recognition"],"materials":["paper","poster","video"]},{"uid":"620","displayId":"LBD-620","title":"A Diagnostic Evaluation Framework for AI-Generated Cover Songs","authors":"Yingxin Liang","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUGRN86","pageUrl":"https://ismir2026program.ismir.net/lbd_620.html","presenterMode":"","keywords":[],"materials":[]},{"uid":"621","displayId":"LBD-621","title":"SheetSage-Pro: Bilingual Lyric-Aligned Lead-Sheet Transcription from Music Audio","authors":"Ping Wang; Fang-Duo Tsai; Wei-Jaw Lee; Yi-Hsuan Yang","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQAQ5H7","pageUrl":"https://ismir2026program.ismir.net/lbd_621.html","presenterMode":"onsite","keywords":["Lead-Sheets","Lyrics","Alignment","Transcription"],"materials":["paper","poster","video"]},{"uid":"622","displayId":"LBD-622","title":"Imaxes Térmicas: Mapping Thermal Imagery to Seeded Open Scores for Networked Improvisation","authors":"Xairo Campos Blanco","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ9KFUL","pageUrl":"https://ismir2026program.ismir.net/lbd_622.html","presenterMode":"online","keywords":["Image-to-music mapping","Interactive music systems","Algorithmic composition","Networked improvisation","Open scores"],"materials":["paper","poster","video"]},{"uid":"624","displayId":"LBD-624","title":"Modulation Type and Amount Shape Human Perception of Timbral Change in Wavetable Synthesis","authors":"Sachin Subramanian; Iran Roman; Johan Pauwels; Joshua Reiss; Christopher Mitcheltree","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJ9562Y","pageUrl":"https://ismir2026program.ismir.net/lbd_624.html","presenterMode":"online","keywords":["modulation perception","listening study","timbre modulation","wavetable synthesis"],"materials":["paper","poster","video"]},{"uid":"626","displayId":"LBD-626","title":"MuSP-Bench: Advanced Multimodal Benchmarking of Music Understanding across Score and Performance","authors":"Milan Liessens Dujardin; Song-Ze Yu; Kevin Miao","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQB302D","pageUrl":"https://ismir2026program.ismir.net/lbd_626.html","presenterMode":"onsite","keywords":["Music understanding","multimodal AI","score and performance","benchmark"],"materials":["paper","poster","video"]},{"uid":"627","displayId":"LBD-627","title":"Hands Tell the Technique: Audio-Visual Playing Technique Classification on CCOM-HuQin","authors":"Jimeng Tian","channelUrl":"https://ismir2026.slack.com/archives/C0CB78U5P32","pageUrl":"https://ismir2026program.ismir.net/lbd_627.html","presenterMode":"online","keywords":["playing technique detection","audio-visual fusion","multimodal MIR","hand pose estimation","erhu"],"materials":["paper","poster","video"]},{"uid":"628","displayId":"LBD-628","title":"LayerIt: Towards a Framework for Time-Aligned, Composable Music Visualizations","authors":"Fernando Azeredo; António Sá Pinto","channelUrl":"https://ismir2026.slack.com/archives/C0CACMTEM6W","pageUrl":"https://ismir2026program.ismir.net/lbd_628.html","presenterMode":"online","keywords":["Visualization","Audio-to-Score Alignment","Python","Beat-Tracking"],"materials":["paper","poster","video"]},{"uid":"629","displayId":"LBD-629","title":"WAVEMULATOR: An Explainable Multi-View Web Prototype for Structural Music Similarity Analysis","authors":"Chetan Teja Rachapudi","channelUrl":"https://ismir2026.slack.com/archives/C0CACMSL9NE","pageUrl":"https://ismir2026program.ismir.net/lbd_629.html","presenterMode":"online","keywords":["Structural Music Similarity","Music Information Retrieval","Explainable MIR","Multi-View Audio Analysis","Music Structure Analysis"],"materials":["paper","poster","video"]},{"uid":"630","displayId":"LBD-630","title":"Kon–Tools: A Symbolic-to-Audio Toolchain for Carnatic Music Percussion","authors":"Kwaaku Boamah-Powers; Huixi Wang; Akshay Anantapadmanabhan; Gopika Krishnan; Carlos Guedes","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUHH65C","pageUrl":"https://ismir2026program.ismir.net/lbd_630.html","presenterMode":"onsite","keywords":["carnatic music","assisted composition","synthetic data"],"materials":["paper","poster","video"]},{"uid":"631","displayId":"LBD-631","title":"The Anantapadmanabhan Collection: An Annotated Dataset for Computational Carnatic Percussion","authors":"Gopika Krishnan; Akshay Anantapadmanabhan; Huixi Wang; Kwaaku Boamah-Powers; André Holzapfel; Kaustuv Kanti Ganguli; Carlos Guedes","channelUrl":"https://ismir2026.slack.com/archives/C0CB78US2LQ","pageUrl":"https://ismir2026program.ismir.net/lbd_631.html","presenterMode":"onsite","keywords":["Carnatic percussion","konnakol","mridangam","dataset","automatic transcription"],"materials":["paper","poster","video"]},{"uid":"632","displayId":"LBD-632","title":"Representational alignment of symbolic and acoustic embeddings and implications for downstream performance","authors":"Prajjwal Bhattarai; Melanie Meng Xue; Margaret Leighton","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQBPERF","pageUrl":"https://ismir2026program.ismir.net/lbd_632.html","presenterMode":"onsite","keywords":["representational alignment","symbolic–acoustic music embeddings","probing"],"materials":["paper","poster","video"]},{"uid":"633","displayId":"LBD-633","title":"The Creative Memory Engine: Realization-Oriented AI for Independent Musicians","authors":"Todd Anstis","channelUrl":"https://ismir2026.slack.com/archives/C0CB78V2L7J","pageUrl":"https://ismir2026program.ismir.net/lbd_633.html","presenterMode":"online","keywords":["music information retrieval","creative AI","songwriting","human-AI co-creation","realization-oriented AI"],"materials":["paper","poster","video"]},{"uid":"636","displayId":"LBD-636","title":"Who Is Conducting? A Conductor Skeleton Dataset and Identification Study from In-the-Wild Orchestral Concert Video","authors":"Jiyun Kim; Dasaem Jeong","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUJ8R2A","pageUrl":"https://ismir2026program.ismir.net/lbd_636.html","presenterMode":"online","keywords":["conducting","performance","video","skeleton","dataset"],"materials":["paper","poster","video"]},{"uid":"637","displayId":"LBD-637","title":"Music Information Retrieval from Sung Arabic: Integrating Language, Prosody, Maqam, and Musical Performance","authors":"Maroun Rahi","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUJ864W","pageUrl":"https://ismir2026program.ismir.net/lbd_637.html","presenterMode":"online","keywords":["Sung Arabic","maqam","language-centric MIR","prosody","multimodal representation"],"materials":["paper","poster","video"]},{"uid":"638","displayId":"LBD-638","title":"Partially Coherent Mixing for Nine-Stem Source Separation of Korean Traditional (Gugak) Ensembles","authors":"WonJae Gye; Dasaem Jeong","channelUrl":"https://ismir2026.slack.com/archives/C0CACMSFZCJ","pageUrl":"https://ismir2026program.ismir.net/lbd_638.html","presenterMode":"online","keywords":["music source separation","non-Western music","Korean traditional music","heterophony","data augmentation"],"materials":["paper","poster","video"]},{"uid":"639","displayId":"LBD-639","title":"A Lightweight Browser-Based Optical Music Recognition Tool for Late-Medieval Plainchant Sources","authors":"Jorge Calvo-Zaragoza; Joan Cerveto-Serrano; Robert Klugseder; Antonio Ríos-Vila; David Rizo; Adrian Rosello","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUH6GPQ","pageUrl":"https://ismir2026program.ismir.net/lbd_639.html","presenterMode":"online","keywords":["Optical Music Recognition","Plainchant","Music Encoding","Symbolic","Handwritten Text Recognition"],"materials":["paper","poster","video"]},{"uid":"640","displayId":"LBD-640","title":"Rethinking Allied Raga Confusion: A Graph-Structured Approach to Carnatic Raga Classification.","authors":"Sragvi Prerana Kalluri; Priyanshu Singh; Vikram Pudi","channelUrl":"https://ismir2026.slack.com/archives/C0CACMUD2LA","pageUrl":"https://ismir2026program.ismir.net/lbd_640.html","presenterMode":"online","keywords":["Video is poster with voice over of agent with subtitles.","Mentions the names of all authors. Never mentions presenter name"],"materials":["paper","poster","video"]},{"uid":"641","displayId":"LBD-641","title":"What Makes Music Hit? Probing Music Embedding Against a DJ's Perceptual Similarity","authors":"Vincent Herranen","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJBHZ8C","pageUrl":"https://ismir2026program.ismir.net/lbd_641.html","presenterMode":"online","keywords":["music similarity","layer-wise probing","self-supervised audio embeddings","single-listener perceptual evaluation"],"materials":["paper","poster","video"]},{"uid":"642","displayId":"LBD-642","title":"TimbreTune","authors":"Ben Cole; Annie Chu","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQBTA4R","pageUrl":"https://ismir2026program.ismir.net/lbd_642.html","presenterMode":"online","keywords":["timbre","voice conversion","daw","vst"],"materials":["paper","poster","video"]},{"uid":"643","displayId":"LBD-643","title":"Playing Along in Time and Key: Causal Rhythmic and Harmonic Grounding for Real-Time Music Accompaniment","authors":"Kevin Bretz; Derya Soydaner; Aske Plaat","channelUrl":"https://ismir2026.slack.com/archives/C0C9XG9SMPZ","pageUrl":"https://ismir2026program.ismir.net/lbd_643.html","presenterMode":"online","keywords":["music","GenAI","real-time","HCI","collaborative"],"materials":["paper","poster","video"]},{"uid":"644","displayId":"LBD-644","title":"Genre-conditioned Key Profile Selection for Untagged Music Catalogs","authors":"Pradeep Sathishkumar","channelUrl":"https://ismir2026.slack.com/archives/C0CB78VK07J","pageUrl":"https://ismir2026program.ismir.net/lbd_644.html","presenterMode":"online","keywords":["key detection","key profiles","genre classification","electronic dance music","music tagging"],"materials":["paper","poster","video"]},{"uid":"645","displayId":"LBD-645","title":"Understanding and Automating Music-Art Mediation: From Behavioural Evidence to AI-Based Selection","authors":"Disha Pawar; Mohammad Anzer Khan; Khushi Daga; Subhrojyoti Roy Chaudhuri; Blessin Varkey; Akash Mohan","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LKSA2X","pageUrl":"https://ismir2026program.ismir.net/lbd_645.html","presenterMode":"online","keywords":["music-art mediation","artwork-aware music selection","cross-modal perception","LLM-based reasoning","museum experience"],"materials":["paper","poster","video"]},{"uid":"646","displayId":"LBD-646","title":"Music Theory Generation from Arbitrary Frequency Sequences Using Large Language Models","authors":"Tatsunori Hirai","channelUrl":"https://ismir2026.slack.com/archives/C0CACMURXAN","pageUrl":"https://ismir2026program.ismir.net/lbd_646.html","presenterMode":"online","keywords":["Music Theory Generation","LLM","Microtonal Music","Interactive system"],"materials":["paper","poster","video"]},{"uid":"647","displayId":"LBD-647","title":"Textbook-Constrained Four-Part Harmonization in a Tablet Notation Editor: A Same-Grader Comparison with Bach and DeepBach","authors":"Bolong Jiang","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJAT68L","pageUrl":"https://ismir2026program.ismir.net/lbd_647.html","presenterMode":"","keywords":[],"materials":[]},{"uid":"649","displayId":"LBD-649","title":"Music Structure-Aware FX Automation from Natural Language","authors":"Dongkyun Kim; Kyogu Lee","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JSQ4FP","pageUrl":"https://ismir2026program.ismir.net/lbd_649.html","presenterMode":"online","keywords":["Audio Effect Automation","Music Post-Production","Time-Varying Audio Processing","Large Language Models"],"materials":["paper","poster","video"]},{"uid":"650","displayId":"LBD-650","title":"Text2Groove: Text-Guided Drum MIDI Humanization","authors":"Adrian Lam; Hidetaka Imamura; Georgios Velissaridis","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JSULN9","pageUrl":"https://ismir2026program.ismir.net/lbd_650.html","presenterMode":"online","keywords":["MIDI humanization","expressive performance modeling","microtiming","natural language conditioning","conditional variational autoencoder"],"materials":["paper","poster","video"]},{"uid":"651","displayId":"LBD-651","title":"Does a Compact Music Encoder Need the Full Teacher Representation?","authors":"Hsieh-Yu Yang; Timothy K. Shih","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JT9ANR","pageUrl":"https://ismir2026program.ismir.net/lbd_651.html","presenterMode":"online","keywords":["Representation Learning","Knowledge Distillation","Model Compression"],"materials":["paper","poster","video"]},{"uid":"652","displayId":"LBD-652","title":"Classifying Irish Tune Types from ABC Notation with a Character-Level LSTM","authors":"Mario Vallejo Reyes","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LMA0SX","pageUrl":"https://ismir2026program.ismir.net/lbd_652.html","presenterMode":"online","keywords":["Irish traditional music","ABC notation","tune-type classification","recurrent neural networks","symbolic music information retrieval"],"materials":["paper","poster","video"]},{"uid":"653","displayId":"LBD-653","title":"TEMPSHIFT: Retrieving Functionally Relevant yet Sonically Distinct Alternatives to Temp Tracks","authors":"Ningyue Yang; Yang Yu","channelUrl":"https://ismir2026.slack.com/archives/C0CACMVRQD8","pageUrl":"https://ismir2026program.ismir.net/lbd_653.html","presenterMode":"onsite","keywords":["Temp‑track retrieval","dramatic function preservation","maximal‑marginal‑relevance reranking","CLAP audio embedding","film‑music prototyping"],"materials":["paper","poster","video"]},{"uid":"654","displayId":"LBD-654","title":"Distorter: A Controllable, Real-Time Harsh-Voice Effect for Speech and Singing","authors":"Beomjin Park; Kyogu Lee","channelUrl":"https://ismir2026.slack.com/archives/C0CB790S86L","pageUrl":"https://ismir2026program.ismir.net/lbd_654.html","presenterMode":"online","keywords":["Harsh voice effects","Extreme metal vocals","Vocal technique conversion","Time-varying intensity control","Real-time audio processing"],"materials":["paper","poster","video"]},{"uid":"655","displayId":"LBD-655","title":"StreamMUSE+: Towards More Stable and Interactive Real-Time Piano Accompaniment Generation","authors":"Lekai Qian; Yuan-Hsin Chen; Jiadong Zhang; Bowen Zheng; Andrew H. Yang; Xiaosong Ma; Ziyu Wang","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUR3D46","pageUrl":"https://ismir2026program.ismir.net/lbd_655.html","presenterMode":"onsite","keywords":["real-time music generation","automatic piano accompaniment","human–AI music interaction","symbolic music generation","streaming inference"],"materials":["paper","poster","video"]},{"uid":"656","displayId":"LBD-656","title":"GEM: Toward Controllable Music Generation for Underrepresented Music Within and Across Traditions","authors":"Yixuan Xiao; Yibei Xiao","channelUrl":"https://ismir2026.slack.com/archives/C0CACN0G9KL","pageUrl":"https://ismir2026program.ismir.net/lbd_656.html","presenterMode":"online","keywords":["symbolic music generation","underrepresented musical traditions","music representation","controllable generation","Transformer"],"materials":["paper","poster","video"]},{"uid":"657","displayId":"LBD-657","title":"La Tumbadora: A Conga Solos Dataset","authors":"Jose Omar Giraldo; Daniel Gómez-Marín","channelUrl":"https://ismir2026.slack.com/archives/C0CACN0K3CJ","pageUrl":"https://ismir2026program.ismir.net/lbd_657.html","presenterMode":"online","keywords":["Hand percussion dataset","Drum transcription","Source Separation"],"materials":["paper","poster","video"]},{"uid":"658","displayId":"LBD-658","title":"CloudEar: Integrating Perceptual, Diagnostic, and Cross-Modal Evidence for Music Evaluation","authors":"Qiqi He; Anqi Huang","channelUrl":"https://ismir2026.slack.com/archives/C0CA7JYHV5K","pageUrl":"https://ismir2026program.ismir.net/lbd_658.html","presenterMode":"onsite","keywords":["Music Evaluation","Text-Music Alignment","Generative Music","Musicality Assessment","Audio Quality Assessment"],"materials":["paper","poster","video"]},{"uid":"659","displayId":"LBD-659","title":"PianoKontext: Expressive Performance Rendering from Deadpan Context","authors":"Dmitrii Gavrilev","channelUrl":"https://ismir2026.slack.com/archives/C0CB7925HBJ","pageUrl":"https://ismir2026program.ismir.net/lbd_659.html","presenterMode":"online","keywords":["expressive performance rendering","music generation","flow matching","classical music"],"materials":["paper","poster","video"]},{"uid":"660","displayId":"LBD-660","title":"LogicProGym: Creating a Shared Musical Environment for Humans and Learning Agents","authors":"Ziyun Liu; Ken Déguernel; Florent Berthaut","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUR4686","pageUrl":"https://ismir2026program.ismir.net/lbd_660.html","presenterMode":"online","keywords":["musical agents","reinforcement learning","human–AI interaction","digital audio workstations","Gymnasium"],"materials":["paper","poster","video"]},{"uid":"661","displayId":"LBD-661","title":"A Simple Real-Time Enharmonic Pitch Speller","authors":"Alexis Doualle","channelUrl":"https://ismir2026.slack.com/archives/C0CA6LSN18B","pageUrl":"https://ismir2026program.ismir.net/lbd_661.html","presenterMode":"onsite","keywords":["pitch spelling","real-time MIDI","Tonnetz","music theory"],"materials":["paper","poster","video"]},{"uid":"662","displayId":"LBD-662","title":"The Shape of Things to Come: Learning Setlist Design from Musical Features","authors":"Anna Kruspe; Mark Gotham","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUSA230","pageUrl":"https://ismir2026program.ismir.net/lbd_662.html","presenterMode":"online","keywords":["setlist","pointer networks","sequence ordering","playlist","live music"],"materials":["paper","poster","video"]},{"uid":"663","displayId":"LBD-663","title":"Natural-Language Playlist Generation in Private DJ Libraries: A Blinded Side-by-Side Demo","authors":"Adin Segall; George Fazekas","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQL50AV","pageUrl":"https://ismir2026program.ismir.net/lbd_663.html","presenterMode":"online","keywords":["music recommendation","playlist generation","natural language","user study","DJ libraries"],"materials":["paper","poster","video"]},{"uid":"664","displayId":"LBD-664","title":"AMIDI: Audio-Supervised Event Modeling for Expressive MIDI Generation","authors":"Qingyang Lyu; Bo Dai","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUW7PE2","pageUrl":"https://ismir2026program.ismir.net/lbd_664.html","presenterMode":"online","keywords":["Expressive MIDI Generation","Symbolic Music Modeling","Audio-Supervised Pretraining","Audio–MIDI Modeling","Musical Performance Expression"],"materials":["paper","poster","video"]},{"uid":"665","displayId":"LBD-665","title":"Modeling What's Fake, Not What's Real: One-Class AI Music Detection","authors":"Jethro Rongyue Lee; Jeffrey Gordon; George Buck; Cedric Lam","channelUrl":"https://ismir2026.slack.com/archives/C0CB7984AN4","pageUrl":"https://ismir2026program.ismir.net/lbd_665.html","presenterMode":"online","keywords":["AI music detection","normalizing flow","artificial intelligence","artifacts","generation"],"materials":["paper","poster","video"]},{"uid":"666","displayId":"LBD-666","title":"Can Notes Alone Be Jazzy?","authors":"Robert Wienroeder","channelUrl":"https://ismir2026.slack.com/archives/C0CAAUXHRF0","pageUrl":"https://ismir2026program.ismir.net/lbd_666.html","presenterMode":"onsite","keywords":["variational autoencoder","latent space steering","jazz","symbolic music","style transfer"],"materials":["paper","poster","video"]},{"uid":"667","displayId":"LBD-667","title":"MixTrace: A Per-Track Measurement Instrument and Chain Formalization for Capturing Human Mixing Trajectories","authors":"Xiaoyun Ma; Nan Nan; Xiaobing Li; Xiaohong Guan","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJNRRDJ","pageUrl":"https://ismir2026program.ismir.net/lbd_667.html","presenterMode":"online","keywords":["music mixing","mixing trajectory","DAW plugin","operation chain formalization","music production analytics"],"materials":["paper","poster","video"]},{"uid":"668","displayId":"LBD-668","title":"Recovering 3D Upper-Body Shapes from Violin Performance Videos","authors":"Le-Ying Chen; Yu-Fen Huang; Li Su","channelUrl":"https://ismir2026.slack.com/archives/C0CB797M2F2","pageUrl":"https://ismir2026program.ismir.net/lbd_668.html","presenterMode":"online","keywords":["Music Performance Analysis","3D Human Motion","Human Mesh Recovery","Violin Performance"],"materials":["paper","poster","video"]},{"uid":"669","displayId":"LBD-669","title":"Aligning Incomplete Lyrics: Local Improvements and Collateral Errors","authors":"Kyounghoon Han","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJPHPGU","pageUrl":"https://ismir2026program.ismir.net/lbd_669.html","presenterMode":"onsite","keywords":["Lyrics alignment","Incomplete lyrics","CTC forced alignment","Wildcard gap handling","Alignment robustness"],"materials":["paper","poster","video"]},{"uid":"670","displayId":"LBD-670","title":"CompMe: A Design Pattern for Human-AI Musical Co-Creation","authors":"Prashanth Thattai Ravikumar","channelUrl":"https://ismir2026.slack.com/archives/C0CA7K4V63X","pageUrl":"https://ismir2026program.ismir.net/lbd_670.html","presenterMode":"onsite","keywords":["Real-Time Music Generation","MIDI Drum Generation","Rhythm modelling","Negotiated control","Human-AI co-creation"],"materials":["paper","poster","video"]},{"uid":"671","displayId":"LBD-671","title":"Natural Language MIDI Loop Retrieval via Structured LLM Translation and Deterministic Scoring","authors":"Nimrod Adar; Hagar Kafri; Sharon Gannot","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQRKJ0Z","pageUrl":"https://ismir2026program.ismir.net/lbd_671.html","presenterMode":"online","keywords":["MIDI loop retrieval","large language models","natural language interfaces","MIDI loop taxonomy","structured LLM retrieval"],"materials":["paper","poster","video"]},{"uid":"672","displayId":"LBD-672","title":"Oh Sheet!: A Browser-Based Optical Music Recognition Tool for Classical and Handwritten Jazz Lead Sheets","authors":"Vansh Chugh; Anh Phan; Khizran Fatima","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQWMXSM","pageUrl":"https://ismir2026program.ismir.net/lbd_672.html","presenterMode":"onsite","keywords":["Optical Music Recognition","Democratization of MIR Research","Software Demo"],"materials":["paper","poster","video"]},{"uid":"673","displayId":"LBD-673","title":"Where Poetry Meets Music: An Interactive System for Synchronized Reading, Listening, and Wearable EEG","authors":"Wentao Yuan; Pedro Sodre; Richard Boulanger","channelUrl":"https://ismir2026.slack.com/archives/C0CAGJTC7EG","pageUrl":"https://ismir2026program.ismir.net/lbd_673.html","presenterMode":"online","keywords":["Brainwave","Music","Muse","EEG","Csound"],"materials":["paper","poster","video"]},{"uid":"674","displayId":"LBD-674","title":"Atlas Flow: Explicit Timbre-Trajectory Modeling for a Navigable Latent-Space Synthesizer","authors":"Jiayi Hu; Yuxuan Shi; Yifeng Huang; Jiangnan Zhang; Jinyue Guo","channelUrl":"https://ismir2026.slack.com/archives/C0CACNC509G","pageUrl":"https://ismir2026program.ismir.net/lbd_674.html","presenterMode":"online","keywords":["Neural audio synthesis","timbre space","flow matching","interactive timbre control","real-time synthesizer"],"materials":["paper","poster","video"]},{"uid":"675","displayId":"LBD-675","title":"Generative Music Tools for Closed DAWs","authors":"Terry Ma; Chris Donahue","channelUrl":"https://ismir2026.slack.com/archives/C0CA6M3LJTD","pageUrl":"https://ismir2026program.ismir.net/lbd_675.html","presenterMode":"online","keywords":["DAW parser","Logic pro x","music continuation"],"materials":["paper","poster","video"]},{"uid":"676","displayId":"LBD-676","title":"Small Supervised Embeddings Beat Zero-Shot Foundation Models on Music Classification Tasks in Essentia","authors":"Arjun Bahuguna; Pablo Alonso-Jiménez; Xavier Lizarraga-Seijas; Dmitry Bogdanov","channelUrl":"https://ismir2026.slack.com/archives/C0CACNDASHY","pageUrl":"https://ismir2026program.ismir.net/lbd_676.html","presenterMode":"online","keywords":["music classification","audio representation learning","foundation models","supervised learning","essentia"],"materials":["paper","poster","video"]},{"uid":"677","displayId":"LBD-677","title":"TimbreCLAP: Adapting CLAP for Timbre-Oriented Sound Retrieval","authors":"Jiayi Hu; Jiangnan Zhang; Yifeng Huang; Yuxuan Shi; Jinyue Guo; Shimeng Li","channelUrl":"https://ismir2026.slack.com/archives/C0CAAV3AD62","pageUrl":"https://ismir2026program.ismir.net/lbd_677.html","presenterMode":"online","keywords":["Timbre-oriented sound retrieval","CLAP fine-tuning","Audio-text contrastive learning","Timbre captioning","Timbre Understanding"],"materials":["paper","poster","video"]},{"uid":"679","displayId":"LBD-679","title":"Neural Codec Latent Distillation for Corpus Retrieval Synthesis on the Elk Stomp","authors":"Trent Eriksen; Jazz Segovia; Dennison Blackett","channelUrl":"https://ismir2026.slack.com/archives/C0CA6M4FYNP","pageUrl":"https://ismir2026program.ismir.net/lbd_679.html","presenterMode":"online","keywords":["knowledge distillation","neural audio codecs","latent granular synthesis","embedded systems"],"materials":["paper","poster","video"]},{"uid":"680","displayId":"LBD-680","title":"Probing Transposition in Audio Embeddings: Human Key Change vs. Algorithmic Pitch-Shift","authors":"Yunying Shu; Ariana Pereira; Connor Zhang","channelUrl":"https://ismir2026.slack.com/archives/C0CB79EG8FJ","pageUrl":"https://ismir2026program.ismir.net/lbd_680.html","presenterMode":"online","keywords":["data augmentation","music representation","pitch-shifting"],"materials":["paper","poster","video"]},{"uid":"681","displayId":"LBD-681","title":"ST.R.E.M (Statistical Recall Music)","authors":"José Kevin Prado Castillo; Jesús Alejandro Rubio González","channelUrl":"https://ismir2026.slack.com/archives/C0CACNDN2JE","pageUrl":"https://ismir2026program.ismir.net/lbd_681.html","presenterMode":"online","keywords":["AI-generated music","Music Information Retrieval","creative lineage","statistical relationship detection","human-in-the-loop systems"],"materials":["paper","poster","video"]},{"uid":"682","displayId":"LBD-682","title":"Text-to-Kit: Set-Level Generation and Evaluation of Playable Drum Kits from Text","authors":"Félix Ménard","channelUrl":"https://ismir2026.slack.com/archives/C0CA7KB5QG5","pageUrl":"https://ismir2026program.ismir.net/lbd_682.html","presenterMode":"online","keywords":["text-to-audio generation","drum sample synthesis","drum kit","set-level evaluation","LoRA fine-tuning"],"materials":["paper","poster","video"]},{"uid":"683","displayId":"LBD-683","title":"Audio2Preset: Reconstructing Editable Synthesizer Presets from Reference Audio","authors":"Letao Sun; Kevin Wang","channelUrl":"https://ismir2026.slack.com/archives/C0CA7KASVPF","pageUrl":"https://ismir2026program.ismir.net/lbd_683.html","presenterMode":"onsite","keywords":["Audio-to-preset","Synthesizer sound matching","Synthesizer parameter estimation","Preset retrieval","Music production"],"materials":["paper","poster","video"]},{"uid":"684","displayId":"LBD-684","title":"Does the Generator Read Its Reference? A Provenance-Audited Arrangement Agent for Ableton Live","authors":"Han Liao","channelUrl":"https://ismir2026.slack.com/archives/C0CAEQXPN5P","pageUrl":"https://ismir2026program.ismir.net/lbd_684.html","presenterMode":"online","keywords":["music arrangement","provenance","human-in-the-loop systems","evaluation of generative music systems","Ableton Live"],"materials":["paper","poster","video"]},{"uid":"685","displayId":"LBD-685","title":"Stressing Source Separation: How Interference and Audio Effects Degrade Vocal Extraction","authors":"Ivana Rasch Chinchilla; Georgia Campbell-Irwin","channelUrl":"https://ismir2026.slack.com/archives/C0CA6M4UC3D","pageUrl":"https://ismir2026program.ismir.net/lbd_685.html","presenterMode":"online","keywords":["Blind source separation (BSS)","interference","vocal extraction","audio effects/processing","BSS evaluation."],"materials":["paper","poster","video"]},{"uid":"686","displayId":"LBD-686","title":"musWM: Roman-Numeral Analysis from Chord-Code Addresses, and a Provenance-Hidden Review Interface","authors":"Oguzhan Tugral","channelUrl":"https://ismir2026.slack.com/archives/C0CACNDGGPL","pageUrl":"https://ismir2026program.ismir.net/lbd_686.html","presenterMode":"","keywords":[],"materials":[]},{"uid":"687","displayId":"LBD-687","title":"Classii: Connecting Music Knowledge for Exploratory Discovery","authors":"Iwao Kobayashi","channelUrl":"https://ismir2026.slack.com/archives/C0C9XGTMK9V","pageUrl":"https://ismir2026program.ismir.net/lbd_687.html","presenterMode":"online","keywords":["Music Information Retrieval","Exploratory Search","Knowledge Graphs","Entity Linking","Music Discovery"],"materials":["paper","poster","video"]},{"uid":"688","displayId":"LBD-688","title":"From Score Ambiguity to Relocation Decisions: Evaluating Intervention Utility in Score Following","authors":"Sungjin Im","channelUrl":"https://ismir2026.slack.com/archives/C0C9XGU7XL7","pageUrl":"https://ismir2026program.ismir.net/lbd_688.html","presenterMode":"onsite","keywords":["score following","music alignment","score ambiguity","decision regret","evaluation methodology"],"materials":["paper","poster","video"]},{"uid":"689","displayId":"LBD-689","title":"MusicScope: Deepfake Music Detection through Latent Contrastive Learning","authors":"Divyansh Rawat; Ishan Pandey; Ramakrishna Sen; M Kiran Reddy","channelUrl":"https://ismir2026.slack.com/archives/C0CA7KBCZK7","pageUrl":"https://ismir2026program.ismir.net/lbd_689.html","presenterMode":"onsite","keywords":["AI-generated music detection","Deepfake detection","Contrastive Learning"],"materials":["paper","poster","video"]}]}