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.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

The paper demonstrates that combining self-forcing and a sequence-level discriminator greatly improves real-time audio accompaniment generation.

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.

This paper can be viewed as a successful follow-up to the SMG system by Wu et al. (the paper entitled "Streaming generation for music accompaniment", uploaded to arXiv in October 2025). The SMG paper nicely defines the problem framework of real-time audio accompaniment generation, but fails to provide a promising solution without impractical look-ahead into future inputs. In contrast, the LiveBand model presented in this submission very nicely addresses the challenges of real-time generation and greatly outperforms its precursor.

The paper is very well-written, clearly explaining the challenges faced and the rationale behind using self-forcing and a sequence-level discriminator to address these challenges. Specifically, the use of self-forcing during training effectively mitigates exposure bias for streaming autoregressive generation, while the sequence-level discriminator ensures long-term temporal and musical coherence. These two methodological components serve as highly valuable, reusable insights for anyone designing low-latency, streaming audio-to-audio generative models.

The experiments are excellently executed, featuring comprehensive objective and subjective evaluations of the proposed model against existing baselines, alongside detailed ablation studies. The inclusion of the "LiveBand_bid" baseline (as a bidirectional upper bound) and "LiveBand_int" (using internal real stems) is smart and provides extra insights. The analysis of the streaming latency is also highly informative.

In general, this paper presents a breakthrough in the challenging and intriguing task of audio generation, with insights that have high potential to influence the way researchers approach this task in the near future. It may present a game-changing solution that marks the beginning of increased research and even commercial deployment of real-time music generation systems, which could fundamentally change the way people interact with music.

In the camera-ready version of the paper, I'd love to see some comments on the performance of the LiveBand_int variant. While there are objective evaluation results in the paper and audio examples on the demo page, the authors did not explicitly discuss its performance in the text.

It seems the authors did not promise to open-source their model. I strongly encourage them to consider doing so.

Q25. Meta-review and final comments for authors (to be completed after the discussion phase)

All reviewers agree that this is an exceptionally strong submission, with two of the reviewers and myself noting that it possesses award-winning quality. It makes a great contribution to the community.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

Streaming generation of musical audio accompaniments conditioned on live audio under various amounts of anticipation.

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.

The paper outlines a method for streaming accompaniment generation in the audio domain. The model design fields the primary fault in previous streaming formulations around exposure bias which at inference can accumulate local errors over the course of time. This is accomplished by means of a sequence-level discriminator which avoids the need for teacher forcing in order to ensure that train / eval scenarios are effectively matched. It uses CodiCodec as its baseline latent and at inference their model constitutes AR over continuous latents.

The results feel significantly better than similar models that were also trained on Slakh. I also appreciate that they specifically underscore the current throughput capabilities under the model to confirm that streamability is possible, as well as some engineering-level considerations.

As such, my verdict is an accept and am looking forward to the contribution this will bring to the ISMIR community.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

This paper elegantly resolves real-time music accompaniment generation by (1) aligning training and inference distributions by design using independent per-step noise as the generator input, and (2) employing a discriminator to ensure holistic sequence coherenc.

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.

The authors present LiveBand, a real-time, strictly causal audio accompaniment generation system. To address the long-term drift and exposure bias commonly found in autoregressive generative models, the authors employ a parallel forward pass during training with sequence-level adversarial supervision (R3GAN) in a continuous latent space. To make the system robust for real-time deployment, they integrate Attention Sinks to manage KV-cache drift and propose an Adaptive Gradient Penalty (AdaGP) to stabilize adversarial training dynamically. Objective and subjective metrics show it outperforms a causal baseline, StreamMusicGen (SMG).

Strengths:

1. Clear Motivation and Practical Framework: The paper clearly identifies exposure bias as the root cause of drift in real-time causal generation. The proposed framework efficiently addresses this by equating the training and inference conditions through single-pass generation and sequence-level supervision. This is a solid engineering design.
2. Comprehensive Engineering Details: The authors systematically address the practical hurdles of their approach. The inclusion of Attention Sinks for long-horizon streaming and the introduction of AdaGP to alleviate the notorious instability of GAN training demonstrate a well-thought-out system implementation.

Weaknesses:

1. Missing comparison in the Demo page: The demo page lacks direct side-by-side audio comparisons using the exact same input mix. Providing controlled comparisons between LiveBand and the baseline (SMG) is crucial for readers to intuitively perceive the claimed improvements in drift and mix adherence.
2. Limited Baseline Comparisons: The paper only compares with SMG. Including SOTA offline models (e.g., FastSAG, AnyAccomp, LaDABand) would better demonstrate the real-time latency vs. audio quality trade-off.

Overall, this is a solid and complete system paper that provides a well-engineered solution to a specific problem (real-time music accompaniment). My recommendation is strong accept.

Q18. Write ONE line (in your own words) with the main take-home message from the paper.

The authors develop a streaming accompaniment model that works with an autoregressive GAN-based training method thus solving the problem of exposure bias with autoregressive models trained via teacher forcing.

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.

Strengths:

• This paper explains well the difficulties behind the task of developing a Live accompaniment generative model (drifting between training and inference for teacher forced models)
• It comes with an interesting way to solve it by using a novel autoregressive GAN-based method.
• To me, it looks like the first model that actually works for this task: the sample page shows that the model manages to predict the chords of the song with no strong conditioning (only the instrument label is fed). We see on some examples that the model self correct itself the same way as an actual musician would do.
• The model has been designed so that the inference can actually be streaming and I would be very interested to see a demo where the model is deployed.

Weaknesses: I don't see any particular weakness apart that it would be interesting to iterate on this project to scale it with more data in order to achieve better quality.

Addition to recent relevant works to the Related Work section. Minimal changes to the formatting.