P7-3: Unified Music Identification for Tracks and Versions
R. Oguz Araz, Joan Serrà, Yuki Mitsufuji, Xavier Serra, Dmitry Bogdanov
Subjects: Indexing and querying ; Open Review ; Alignment, synchronization, and score following ; Applications ; Similarity metrics ; Representations of music ; Music retrieval systems ; Musical features and properties ; Analysis tasks ; Pattern matching and detection
Presented Remote
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.
Q18. Write ONE line (in your own words) with the main take-home message from the paper.
This manuscript compares track identification and version identification and intends to provide a unified task description (and potentially a unified solution).
Q23. Main review and comments for the authors (to be completed before the discussion phase). Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.
In general, this paper is well organised and well written. The intention of the author is clear. The readability of the manuscript could be significantly improved if more background information will be give such as the difference between track and version (with more detailed description and examples); and the potential (different) applications of VI and TI.
Q25. Meta-review and final comments for authors (to be completed after the discussion phase)
This paper is general well structured. All reviewers recommend that the paper be accepted. However, there are several minor issues found. The author should carefully follow the suggestions to improve the manuscript.
Q18. Write ONE line (in your own words) with the main take-home message from the paper.
VI models can handle TI reasonably well, but fully subsuming TI within VI remains an open challenge due to retrieval-level constraints rather than fundamental limitations of the embedding space.
Q23. Main review and comments for the authors. Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.
This paper investigates whether VI can subsume TI within a unified framework. The authors propose a benchmark evaluating both tasks across three query conditions, clean, manipulated, and manipulated and degraded, and compare seven state-of-the-art models. They additionally train a baseline model, Fish, that targets both tasks simultaneously. The main finding is that VI models exhibit meaningful TI capability, while TI models show essentially no VI capability. The paper further attributes Fish's lower TI performance to retrieval-level constraints rather than limitations of the embedding space.
Strengths (1) The evaluation scale, with over 633k TI queries and 372k VI queries, is much larger than in previous studies. The modeling of the music editing-playback-capture chain and the reporting of 95% confidence intervals also reflect a carefully designed experimental framework.
(2) The idea of positioning VI and TI as potentially unified tasks is a novel perspective, as these two communities have largely developed independently of each other. The proposed benchmark, which evaluates both tasks on a shared database, is a valuable contribution to the field.
(3) I found the analysis in Tables 3 and 4, which separates the effects of boundary misalignment and information mismatch, to be one of the most insightful parts of the paper. It helpfully shows that Fish's lower TI performance is caused by retrieval design constraints rather than a fundamental limitation of the embedding space.
(4) I appreciate the authors' commitment to releasing the code, metadata, and model weights, as this will help other researchers build on this work and will increase the long-term utility of the benchmark.
Weaknesses (1) The paper proposes that a single system should handle both TI and VI, but I think the justification for this is not yet convincing. In practice, these two tasks serve different purposes: TI is needed to identify the exact recording, as in broadcast monitoring or copyright enforcement, while VI is needed to find all versions of a song across different performances. These goals can conflict, and a system that groups a cover song with the original may be useful for music discovery but problematic in rights management contexts. I would encourage the authors to discuss more carefully under what conditions a unified system is preferable to two separate systems, and what trade-offs users should expect.
(2) The paper claims that VI models can handle TI if retrieval constraints are addressed, but this is only shown for Fish. The same analysis is not carried out for CLEWS or other long-context VI models. To make this a general claim rather than a finding specific to one model, the authors would need to show that the same factors, boundary misalignment and information mismatch, explain the TI performance gap across multiple VI models.
(3) The paper does not have a proper related work section. Section 2 describes existing models but does not clearly explain how this paper differs from or improves upon prior work.
(4) Table 6 tests different versions of the training data pipeline. However, the paper does not test the effect of the architectural changes made to Fish, specifically the removal of the frequency stride and the addition of L2 normalization. Without these experiments, it is unclear which design choices actually drive Fish's performance.
(5) The paper states that code, metadata, and model weights will be released. However, several details needed to reproduce the results from the paper text alone are missing:
- The parameters used to compute the CQT, such as hop size and frequency range, are not reported.
- The random seed used to sample query segments is not provided, so the exact same query set cannot be reconstructed.
- The threshold for grouping local minima into segment-level cliques based on temporal overlap is not specified.
Q18. Write ONE line (in your own words) with the main take-home message from the paper.
The proposed evaluation framework, covering both version identification and track identification as well as their respective augmentations, provides a useful foundation for considering the two tasks jointly.
Q23. Main review and comments for the authors. Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.
This paper proposes a methodology for jointly evaluating version identification and track identification, two tasks that have traditionally been studied independently. In addition, the authors assess the robustness of each task under various augmentation settings. The results in Tables 1 and 2 show that the prior work CLEWS achieves strong performance across both tasks. The paper further introduces Fish, a model that adopts an architecture similar to CoverHunter while incorporating improved training methodologies and retrieval stage optimizations. Experimental results suggest that Fish achieves performance comparable to CLEWS on both tasks and demonstrates better robustness under augmentation settings.
Overall, the paper presents a novel evaluation framework that brings together version identification and track identification within a unified setting. However, the description of the baseline model lacks sufficient detail, which raises concerns regarding reproducibility and the ability of future researchers to fully understand and replicate the proposed approach.
Q18. Write ONE line (in your own words) with the main take-home message from the paper.
VI models can already perform TI to some degree, but fully subsuming TI requires overcoming the granularity barrier of boundary misalignment and information mismatch.
Q23. Main review and comments for the authors. Please summarize strengths and weaknesses of the paper. It is essential that you justify the reason for the overall evaluation score in detail. Keep in mind that belittling or sarcastic comments are not appropriate.
・Strength S1. This paper brings studies about TI and VI together into one question, "can VI subsume TI?", and evaluates them on a shared database. This framing is interesting and novel. S2. This paper uses 3 databases, over 1M queries, and 3 conditions (clean, manipulated, manipulated & degraded), and reports a 95% confidence interval for every metric. It also defines the editing-playback-capture chain so that it can be reproduced with audiomentations, and the code, metadata, and model weights will be shared. So other researchers can use it directly. S3. The paper does not just say Fish has lower TI performance than NMFP. It shows that the gap comes from two factors, boundary misalignment and information mismatch. This gives an understanding of why unification is hard, and offers a good guideline for designing unified models in the future.
・Weakness W1. The abstract and conclusion read as if Fish performs well on both tasks. But in Table 1, for clean HR@1, Fish (86.2 / 83.4) is more than 10 points below NMFP (96.9 / 97.4). This does not support the claim that VI can subsume TI. The authors explain this in Section 5.1 as a retrieval constraint rather than an embedding space problem, but the "misalignment = 0" setting in Table 4 is an oracle condition that does not hold in real use, so it is not a realistic counter-argument. Even in Table 5, the clean HR@1 with 10s segments stays at 91.78, still below NMFP. W2. The explanation of the training data construction is too compressed. The clique construction using flood-fill, basin, and hole (Section 4.1) is the core of this paper, but it is hard to follow on a first read. W3. The paper lists code, metadata, and model weights as a shared resource (footnote 1), and this is a major selling point for reproducibility. However, the link is currently broken and the code cannot be accessed. This must be fixed before camera-ready.
Overall, this paper is well-structured and makes a clear contribution. From the perspectives of reproducibility and readability, I strongly recommend making several revisions for the camera-ready version.
Title is simplified. Writing improved following the reviews: the tone of the paper is homogenized across the abstract, results, and conclusion. This is achieved through narrowing the scope.