Musicians and nonmusicians alike use rhythmic sound gestures, such as tapping and beatboxing, to express drum patterns. While these gestures effectively communicate musical ideas, realizing these ideas as fully-produced drum recordings can be time-consuming, potentially disrupting many creative workflows. To bridge this gap, we present TRIA (The Rhythm In Anything), a masked transformer model for mapping rhythmic sound gestures to high-fidelity drum recordings. Given an audio prompt of the desired rhythmic pattern and a second prompt to represent drumkit timbre, TRIA produces audio of a drumkit playing the desired rhythm (with appropriate elaborations) in the desired timbre. Subjective and objective evaluations show that a TRIA model trained on less than 10 hours of publicly-available drum data can generate high-quality, faithful realizations of sound gestures across a wide range of timbres in a zero-shot manner.

Q2 ( I am an expert on the topic of the paper.)

Strongly agree

Q3 ( The title and abstract reflect the content of the paper.)

Agree

Q4 (The paper discusses, cites and compares with all relevant related work.)

Disagree

Q5 ( Please justify the previous choice (Required if “Strongly Disagree” or “Disagree” is chosen, otherwise write "n/a"))

The citations for KAD aren't complete - KAD was already used as "KID" in Nistal et al., Comparing representations for audio synthesis using generative adversarial networks, 27th European Signal Processing Conference (EUSIPCO), 2019

The authors missed several related works regarding drum sample and drum pattern generation. As that is the main topic of the paper, please include such works as well (see below for suggestions)

Drum sample generation:

J. Nistal, S. Lattner, and G. Richard, “DrumGAN: Synthesis of drum sounds with timbral feature conditioning using generative adversarial networks,” in Proc. Int. Soc. Music Inf. Retrieval Conf. (ISMIR), Montréal, Canada, 2020.

A. Lavault, A. Roebel, and M. Voiry, “StyleWaveGAN: Style-based synthesis of drum sounds with extensive controls using generative adversarial networks,” in Proceedings of the Sound and Music Computing Conference (SMC), Saint-Étienne, France, 2022.

J. Drysdale, M. Tomczak, and J. Hockman, “Style-based drum synthesis with GAN inversion,” in Extended Abstracts for the Late-Breaking Demo Session of the 22nd International Society for Music Information Retrieval Conference (ISMIR), Online, 2021

Drum pattern/rhythm generation (audio and symbolic):

G. Alain, M. Chevalier-Boisvert, F. Osterrath, and R. Piché-Taillefer, “DeepDrummer: Generating drum loops using deep learning and a human in the loop,” arXiv preprint arXiv:2008.04391, 2020.

I.-C. Wei, C.-W. Wu, and L. Su, “Generating structured drum patterns using variational autoencoder and self-similarity matrix,” in Proc. Int. Soc. Music Inf. Retrieval Conf. (ISMIR), 2019.

D. P. W. Ellis and J. Arroyo, “Eigenrhythms: Drum pattern basis sets for classification and generation,” in Proc. Int. Soc. Music Inf. Retrieval Conf. (ISMIR), 2004.

D. Gómez-Marín, S. Jordà, and P. Herrera, “Network representations of drum sequences for classification and generation,” Frontiers in Computer Science, vol. 6, 2024. doi: 10.3389/fcomp.2024.1476996

S. Lattner and M. Grachten, “DrumNet: High-level control of drum track generation using learned patterns of rhythmic interaction,” in Proc. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2019. doi: 10.1109/WASPAA.2019.8937229

Q6 (Readability and paper organization: The writing and language are clear and structured in a logical manner.)

Strongly agree

Q7 (The paper adheres to ISMIR 2025 submission guidelines (uses the ISMIR 2025 template, has at most 6 pages of technical content followed by “n” pages of references or ethical considerations, references are well formatted). If you selected “No”, please explain the issue in your comments.)

Yes

Q8 (Relevance of the topic to ISMIR: The topic of the paper is relevant to the ISMIR community. Note that submissions of novel music-related topics, tasks, and applications are highly encouraged. If you think that the paper has merit but does not exactly match the topics of ISMIR, please do not simply reject the paper but instead communicate this to the Program Committee Chairs. Please do not penalize the paper when the proposed method can also be applied to non-music domains if it is shown to be useful in music domains.)

Strongly agree

Q9 (Scholarly/scientific quality: The content is scientifically correct.)

Disagree

Q10 (Please justify the previous choice (Required if “Strongly Disagree” or “Disagree” is chose, otherwise write "n/a"))

• The interpretation of FAD is incorrect

Q11 (Novelty of the paper: The paper provides novel methods, applications, findings or results. Please do not narrowly view "novelty" as only new methods or theories. Papers proposing novel musical applications of existing methods from other research fields are considered novel at ISMIR conferences.)

Agree

Q12 (The paper provides all the necessary details or material to reproduce the results described in the paper. Keep in mind that ISMIR respects the diversity of academic disciplines, backgrounds, and approaches. Although ISMIR has a tradition of publishing open datasets and open-source projects to enhance the scientific reproducibility, ISMIR accepts submissions using proprietary datasets and implementations that are not sharable. Please do not simply reject the paper when proprietary datasets or implementations are used.)

Agree

Q13 (Pioneering proposals: This paper proposes a novel topic, task or application. Since this is intended to encourage brave new ideas and challenges, papers rated “Strongly Agree” and “Agree” can be highlighted, but please do not penalize papers rated “Disagree” or “Strongly Disagree”. Keep in mind that it is often difficult to provide baseline comparisons for novel topics, tasks, or applications. If you think that the novelty is high but the evaluation is weak, please do not simply reject the paper but carefully assess the value of the paper for the community.)

Agree (Novel topic, task, or application)

Q14 (Reusable insights: The paper provides reusable insights (i.e. the capacity to gain an accurate and deep understanding). Such insights may go beyond the scope of the paper, domain or application, in order to build up consistent knowledge across the MIR community.)

Agree

Q15 (Please explain your assessment of reusable insights in the paper.)

I consider the way timbre information is injected as context a reusable insight.

Q16 ( Write ONE line (in your own words) with the main take-home message from the paper.)

Masked language modeling can be used for timbre transfer.

Q17 (This paper is of award-winning quality.)

No

Q19 (Potential to generate discourse: The paper will generate discourse at the ISMIR conference or have a large influence/impact on the future of the ISMIR community.)

Disagree

Q20 (Overall evaluation (to be completed before the discussion phase): Please first evaluate before the discussion phase. Keep in mind that minor flaws can be corrected, and should not be a reason to reject a paper. Please familiarize yourself with the reviewer guidelines at https://ismir.net/reviewer-guidelines.)

Weak accept

Q21 (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.)

The paper introduces a masked Transformer model that produces audio clips of drums based on a timbre and a rhythm prompt. During training, a chosen "buffer" is partially masked, and the masked tokens are predicted based on the unmasked parts. At inference, the timbre prompt is appended to a fully masked segment that contains rhythm features.

The paper is well structured and easy to follow. The sound examples are convincing, and most evaluations are reasonable. There is a major problem with the interpretation of KAD and some minor problems.

Positive: - Well-structured and easy to follow paper with convincing results. - I appreciate the subjective study that shows that both models produce convincing output. It is understood that it wouldn't be fair to ask about timbre, as the MelodyFlow model isn't designed to adhere to a given timbre audio reference.

Questions/Remarks: - The interpretation of KAD is incorrect: "Finally, as shown in Table 3, TRIA produces more realistic drum audio than MelodyFlow on average." The reference set chosen for KAD calculation is from the same distribution as the dataset used for timbre conditioning (i.e., MoisesDB dataset). Therefore it is understandable that the KAD from TRIA is lower than that from MelodyFlow, as MelodyFlow was prompted with random prompts that don't follow the reference data distribution. This test doesn't show "realistic drum audio"; it is rather redundant to the results in Table 2 (timbre column), where understandably, there are no results for MelodyFlow. Similar to omitting results for MelodyFlow in Table 2, the KAD comparison should also be considered invalid. Please remove or clarify that in the camera-ready version. - Why is the buffer chosen anywhere in the sequence during training, if at inference time, the generated part is only at the end? With enough training this doesn't really matter, but it seems to complicate things unnecessarily. Please clarify. - No reason is given why the range [0, 0.1, 0.2] was chosen for the re-noising parameter in MelodyFlow. - No mentioning about the used model architecture in either Abstract or Introduction. Please mention "Masked Transfomer" both in the Abstract and Introduction.

Q22 (Final recommendation (to be completed after the discussion phase) Please give a final recommendation after the discussion phase. In the final recommendation, please do not simply average the scores of the reviewers. Note that the number of recommendation options for reviewers is different from the number of options here. We encourage you to take a stand, and preferably avoid “weak accepts” or “weak rejects” if possible.)

Accept

Q23 (Meta-review and final comments for authors (to be completed after the discussion phase))

Overall, the reviews are positive, with two reviewers assigning Strong Accept and one Weak Accept. The reviewers consistently noted technical novelty, clarity of exposition, and practical relevance of the work. Highlights include:

R1: “This is a very accomplished, polished piece of work which makes a valuable contribution,” and, “clear utility as a musical tool.”

R2: “Very interesting disentangling method… clear contribution toward the emerging task of audio drum pattern synthesis.”

R3: “Strong empirical evidence… modular conditioning and adaptive rhythm representation extendable to other tasks.”

Areas for improvement included:

R1: Clarification of the timbre/rhythm terminology and deeper ablations (e.g., on CFG).

R2: Justification of choices such as the use of DAC codec and ChatGPT-based descriptors; concerns about potential western bias and prompt modality mismatches.

R3: Correction of KAD interpretation and improved clarity on architectural description and training-inference masking mismatch.

Meta-reviewer: Missing references to related work (especially KID/KAD, and prior work in drum synthesis and pattern generation); incorrect interpretation of KAD; lack of related literature on drum pattern/sound generation.

Despite these concerns, the methodological contribution is considered solid and the experimental evaluation largely convincing.

Recommendation The paper is technically sound, clearly written, and introduces a method that is novel and relevant to the ISMIR community. Remaining concerns are largely correctable in the camera-ready version. The authors are encouraged to revise the KAD interpretation, clarify architectural details in the introduction/abstract, and address the missing citations listed in the meta-review. Given the overall strong reviews and technical contribution, I recommend Accept.

Congratulations to the authors for a strong and impactful paper.

Q2 ( I am an expert on the topic of the paper.)

Strongly agree

Q3 (The title and abstract reflect the content of the paper.)

Strongly agree

Q4 (The paper discusses, cites and compares with all relevant related work)

Strongly agree

Q6 (Readability and paper organization: The writing and language are clear and structured in a logical manner.)

Strongly agree

Q7 (The paper adheres to ISMIR 2025 submission guidelines (uses the ISMIR 2025 template, has at most 6 pages of technical content followed by “n” pages of references or ethical considerations, references are well formatted). If you selected “No”, please explain the issue in your comments.)

Yes

Q8 (Relevance of the topic to ISMIR: The topic of the paper is relevant to the ISMIR community. Note that submissions of novel music-related topics, tasks, and applications are highly encouraged. If you think that the paper has merit but does not exactly match the topics of ISMIR, please do not simply reject the paper but instead communicate this to the Program Committee Chairs. Please do not penalize the paper when the proposed method can also be applied to non-music domains if it is shown to be useful in music domains.)

Strongly agree

Q9 (Scholarly/scientific quality: The content is scientifically correct.)

Strongly agree

Q11 (Novelty of the paper: The paper provides novel methods, applications, findings or results. Please do not narrowly view "novelty" as only new methods or theories. Papers proposing novel musical applications of existing methods from other research fields are considered novel at ISMIR conferences.)

Agree

Q12 (The paper provides all the necessary details or material to reproduce the results described in the paper. Keep in mind that ISMIR respects the diversity of academic disciplines, backgrounds, and approaches. Although ISMIR has a tradition of publishing open datasets and open-source projects to enhance the scientific reproducibility, ISMIR accepts submissions using proprietary datasets and implementations that are not sharable. Please do not simply reject the paper when proprietary datasets or implementations are used.)

Strongly agree

Q13 (Pioneering proposals: This paper proposes a novel topic, task or application. Since this is intended to encourage brave new ideas and challenges, papers rated "Strongly Agree" and "Agree" can be highlighted, but please do not penalize papers rated "Disagree" or "Strongly Disagree". Keep in mind that it is often difficult to provide baseline comparisons for novel topics, tasks, or applications. If you think that the novelty is high but the evaluation is weak, please do not simply reject the paper but carefully assess the value of the paper for the community.)

Agree (Novel topic, task, or application)

Q14 (Reusable insights: The paper provides reusable insights (i.e. the capacity to gain an accurate and deep understanding). Such insights may go beyond the scope of the paper, domain or application, in order to build up consistent knowledge across the MIR community.)

Strongly agree

Q15 (Please explain your assessment of reusable insights in the paper.)

My main take-away here is that the proposed setup appears to result in impressive disentanglement between timbre and rhythm conditioning. This seems like a result that could be quite influential in the design of controllable generative audio systems.

Q16 (Write ONE line (in your own words) with the main take-home message from the paper.)

Masked generative modelling of DAC tokens conditioned on a dualized rhythm representation allows disentangled control over timbre and rhythm for realistic drum audio generation.

Q17 (Would you recommend this paper for an award?)

No

Q19 (Potential to generate discourse: The paper will generate discourse at the ISMIR conference or have a large influence/impact on the future of the ISMIR community.)

Agree

Q20 (Overall evaluation: Keep in mind that minor flaws can be corrected, and should not be a reason to reject a paper. Please familiarize yourself with the reviewer guidelines at https://ismir.net/reviewer-guidelines)

Strong accept

Q21 (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.)

I note that anonymity has been somewhat breached here: the authors shared this work widely online at the end of 2024 and I suspect it would be hard to find a reviewer with the appropriate expertise who is not already aware of this contribution. Nonetheless, I have endeavoured to evaluate the paper as impartially as I can.

Overview: The work presents an audio generative model for percussion audio which is designed to disentangle rhythm and timbre conditioning signals. The model builds on VampNet to predict DAC tokens via masked generative modelling in a coarse-to-fine manner. It uses a “dualized” spectral representation with an adaptive crossover frequency as rhythm conditioning, and a DAC token prefix to represent timbre. The result is a model that effectively disentangles drum rhythms from drum timbre, allowing for a mapping from arbitrary recordings of “sound gestures” to realistic drum audio.

Impression: My broad impression is extremely positive. This is a very accomplished, polished piece of work which makes a valuable contribution to an interesting research direction. Further, it continues in a trend of “musician-positive” work on generative AI which I feel sets a constructive example to the ISMIR community. Audio examples demonstrate that the resulting model clearly works well, and has clear utility as a musical tool.

I do have some concerns about the KAD evaluation, with respect to the MelodyFlow baseline. These are detailed below. However, even in light of this, I believe the work is of sufficient value to the community to warrant acceptance.

Some general observations: I wonder about the choice of “rhythm” and “timbre” as terminology for the conditioning signals. These should perhaps be referred to as “gesture” and “sound palette” signals, as the former captures extra-rhythmic aspects such as intensity and coarsely banded spectral centroid, while the latter is really capable of describing multiple timbres of which some or all may be present in the resulting signal, as well as other related timbres which are not captured in the conditioning.

Given the predictability of many drum rhythms and the extreme information bottleneck in the rhythm representation, it’s something of a surprise that the disentanglement was so effective and the model so effectively ignores rhythmic content in the timbre prefix. Of course, the effects-based augmentation probably helped here, but I wonder whether the use of classifier-free guidance played a more important role. If the authors are planning any follow-up work, my opinion is that an ablation of CFG and exploration of sampling weights would be very helpful to readers in understanding how the proposed method actually works. Similarly, it would be helpful to know whether choices like the quantisation of the dualised representation are truly necessary.

Excellent consideration of ethical implications. Perhaps there is a slight gap in the consideration of western bias: it is suggested that the model would function equally well if trained on percussion recordings from different musical cultures, but it does seem that the “dualized” adaptive rhythm representation exhibits a subtle bias to western rhythmic phraseology. All the prior work on this representation appears to focus on drum kit in western popular styles. Regardless, this does not detract from the work as presented, and is simply offered as a consideration for future research.

Criticisms: My main criticism is that the KAD comparison to the MelodyFlow baseline is not under apples-to-apples conditions. TRIA receives timbre conditioning via a prefix of DAC tokens computed from MoisesDB recordings, while MelodyFlow receives a ChatGPT generated text description. I feel this distinction needs to be better emphasised in the text to make clear that MelodyFlow results can not be directly compared to TRIA results.

I question the value of motivating this work on the basis, given in the introduction, that: “to realize a sound gesture as a fully-produced drum arrangement often requires significant time and skill”. I feel this line of reasoning risks alienating musicians by suggesting that it wishes to replace their labour rather than augment their abilities. It also seems a bit contradictory in light of the statement in section 6.1 that the authors view this work “as a means to provide music creators with additional agency”.

To be clear, I'm not challenging the value of the work, just the framing of this particular motivation. Honestly, I feel the affordances of the proposed work stand on their own: it enables multiple new avenues for creative expression, which do not require justification as a mere time-saving contrivance.

I found myself wondering about the performance of TRIA in the face of extreme OOD samples. To me, this would appear to be a very valuable creative use case: i.e. can I “play” any arbitrary sound as a percussion instrument by simply tapping or beatboxing? This is briefly addressed in the online supplement and in Fig. 4, but it would be nice to offer a more concrete evaluation of the model’s behaviour in the face of such extreme timbre conditioning. Honestly, I think such experimentation may have offered more insight into the musical value of the model than the FrameRNN transcription experiment.

I also think this would be particularly revealing w.r.t. the MelodyFlow baseline. The paper points out that TRIA outperforms MelodyFlow despite being much smaller and trained with less data, but this is relatively unsurprising: task-specific models often outperform generalists on in-domain evaluations. I suspect that probing generalisation may reveal some (understandable) limitations of TRIA in comparison to MelodyFlow.

Minutiae: Typo: Tables 1-3 all list MelodyFlow as MelodyFlowFlow

Q2 ( I am an expert on the topic of the paper.)

Agree

Q3 (The title and abstract reflect the content of the paper.)

Agree

Q4 (The paper discusses, cites and compares with all relevant related work)

Strongly agree

Q6 (Readability and paper organization: The writing and language are clear and structured in a logical manner.)

Strongly agree

Q7 (The paper adheres to ISMIR 2025 submission guidelines (uses the ISMIR 2025 template, has at most 6 pages of technical content followed by “n” pages of references or ethical considerations, references are well formatted). If you selected “No”, please explain the issue in your comments.)

Yes

Q8 (Relevance of the topic to ISMIR: The topic of the paper is relevant to the ISMIR community. Note that submissions of novel music-related topics, tasks, and applications are highly encouraged. If you think that the paper has merit but does not exactly match the topics of ISMIR, please do not simply reject the paper but instead communicate this to the Program Committee Chairs. Please do not penalize the paper when the proposed method can also be applied to non-music domains if it is shown to be useful in music domains.)

Strongly agree

Q9 (Scholarly/scientific quality: The content is scientifically correct.)

Strongly agree

Q11 (Novelty of the paper: The paper provides novel methods, applications, findings or results. Please do not narrowly view "novelty" as only new methods or theories. Papers proposing novel musical applications of existing methods from other research fields are considered novel at ISMIR conferences.)

Strongly agree

Q12 (The paper provides all the necessary details or material to reproduce the results described in the paper. Keep in mind that ISMIR respects the diversity of academic disciplines, backgrounds, and approaches. Although ISMIR has a tradition of publishing open datasets and open-source projects to enhance the scientific reproducibility, ISMIR accepts submissions using proprietary datasets and implementations that are not sharable. Please do not simply reject the paper when proprietary datasets or implementations are used.)

Strongly agree

Q13 (Pioneering proposals: This paper proposes a novel topic, task or application. Since this is intended to encourage brave new ideas and challenges, papers rated "Strongly Agree" and "Agree" can be highlighted, but please do not penalize papers rated "Disagree" or "Strongly Disagree". Keep in mind that it is often difficult to provide baseline comparisons for novel topics, tasks, or applications. If you think that the novelty is high but the evaluation is weak, please do not simply reject the paper but carefully assess the value of the paper for the community.)

Agree (Novel topic, task, or application)

Q14 (Reusable insights: The paper provides reusable insights (i.e. the capacity to gain an accurate and deep understanding). Such insights may go beyond the scope of the paper, domain or application, in order to build up consistent knowledge across the MIR community.)

Strongly agree

Q15 (Please explain your assessment of reusable insights in the paper.)

The disentangling method to separate pattern and timber information from audio drum patterns is a reusable insights of this paper. Also the new task of audio drum pattern synthesis (without the intermediary symbolic representation).

Q16 (Write ONE line (in your own words) with the main take-home message from the paper.)

This paper presents a methods to disentangle pattern and timber information from an audio drum pattern for the task of Audio Pattern Generation with timber controlled with text-prompt or audio-prompt and pattern controlled by beat-boxing or tapping

Q17 (Would you recommend this paper for an award?)

No

Q19 (Potential to generate discourse: The paper will generate discourse at the ISMIR conference or have a large influence/impact on the future of the ISMIR community.)

Strongly agree

Q20 (Overall evaluation: Keep in mind that minor flaws can be corrected, and should not be a reason to reject a paper. Please familiarize yourself with the reviewer guidelines at https://ismir.net/reviewer-guidelines)

Weak accept

Q21 (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 starts with the sentence "Musicians and non-musicians alike use rhythmic sound gestures, such as tapping and beat-boxing, to express drum patterns." but doesn't quote any reference to justify this statement. Later on, TRIA is presented as a solution that "produces audio of a drumkit playing the desired rhythm". It would be interesting to have further justification toward why producing an audio recording directly is useful to musicians (and non-musicians), considering that the producing workflow of musicians for example in EDM typically involve some sort of symbolic notation within a DAW or Drum Machines. More interestingly, it would be interesting to justify the other way around: what musicians and non-musicians are doing that requires to produce audio drum pattern. Regarding the technical aspect, it would also be interesting to justify further the use of the Descript Audio Codec to encode highly percussive audio content. Was it compared to other encoders or does the choice rely on another study? The Dualized Rhythm Representation is a very interesting part of this paper and it would be very useful to have some sort of graphic representation of the rhythm feature representation to visualize and insist further on one of the main contribution of the paper. Regarding the experiment, it was mentioned that the descriptions of the drum kits timbers was made with ChatGPT. It would be useful to explain why using a combination of MIR algorithms from the community couldn't be used at this point, or why already existing labeled drum content couldn't be used. In the subjective evaluation, TRIA2Band and MelodyFlow0.2 were selected. Would it be possible to justify this decision? Also since all the listeners are US speakers, it would be interesting to comment on the potential bias resulting of this decision/constraint. Finally, the Ethics Statement is interesting, but I do not know if all those details belongs to this paper. Maybe a separate paper about the ethical aspect of such research could be useful to address some details mentioned in this paragraph. Side comment: mentioning the energy cost of the research without the associated carbon footprint isn't very useful as different countries can have very different kWh/CO2 yearly average ratio.

Q2 ( I am an expert on the topic of the paper.)

Agree

Q3 (The title and abstract reflect the content of the paper.)

Agree

Q4 (The paper discusses, cites and compares with all relevant related work)

Agree

Q6 (Readability and paper organization: The writing and language are clear and structured in a logical manner.)

Agree

Q7 (The paper adheres to ISMIR 2025 submission guidelines (uses the ISMIR 2025 template, has at most 6 pages of technical content followed by “n” pages of references or ethical considerations, references are well formatted). If you selected “No”, please explain the issue in your comments.)

Yes

Q8 (Relevance of the topic to ISMIR: The topic of the paper is relevant to the ISMIR community. Note that submissions of novel music-related topics, tasks, and applications are highly encouraged. If you think that the paper has merit but does not exactly match the topics of ISMIR, please do not simply reject the paper but instead communicate this to the Program Committee Chairs. Please do not penalize the paper when the proposed method can also be applied to non-music domains if it is shown to be useful in music domains.)

Strongly agree

Q9 (Scholarly/scientific quality: The content is scientifically correct.)

Agree

Q11 (Novelty of the paper: The paper provides novel methods, applications, findings or results. Please do not narrowly view "novelty" as only new methods or theories. Papers proposing novel musical applications of existing methods from other research fields are considered novel at ISMIR conferences.)

Agree

Q12 (The paper provides all the necessary details or material to reproduce the results described in the paper. Keep in mind that ISMIR respects the diversity of academic disciplines, backgrounds, and approaches. Although ISMIR has a tradition of publishing open datasets and open-source projects to enhance the scientific reproducibility, ISMIR accepts submissions using proprietary datasets and implementations that are not sharable. Please do not simply reject the paper when proprietary datasets or implementations are used.)

Agree

Q13 (Pioneering proposals: This paper proposes a novel topic, task or application. Since this is intended to encourage brave new ideas and challenges, papers rated "Strongly Agree" and "Agree" can be highlighted, but please do not penalize papers rated "Disagree" or "Strongly Disagree". Keep in mind that it is often difficult to provide baseline comparisons for novel topics, tasks, or applications. If you think that the novelty is high but the evaluation is weak, please do not simply reject the paper but carefully assess the value of the paper for the community.)

Agree (Novel topic, task, or application)

Q14 (Reusable insights: The paper provides reusable insights (i.e. the capacity to gain an accurate and deep understanding). Such insights may go beyond the scope of the paper, domain or application, in order to build up consistent knowledge across the MIR community.)

Agree

Q15 (Please explain your assessment of reusable insights in the paper.)

This paper offers highly reusable insights because it distills general design principles extendable to diverse audio-generation tasks: its modular dual-prompt conditioning cleanly decouples “control signal” and “style example,” enabling straightforward transfer to groove transfer, Foley replacement, or cross-instrument synthesis; its adaptive two-band dualized-rhythm representation encodes temporal accents directly from waveforms, eliminating MIDI/onset prerequisites and thus remaining domain-agnostic; its data-efficient masked-token training demonstrates state-of-the-art quality with <10 h of public data on a 43 M-parameter model, furnishing a concrete benchmark for low-resource scenarios; and its thorough ablations (band splits, masking ratios, guidance weights) plus a commitment to open-sourcing code and checkpoints transform the work from a one-off system into an instructive, transparent blueprint that future researchers can readily reproduce, understand, and adapt beyond the beatbox-to-drums task.

Q16 (Write ONE line (in your own words) with the main take-home message from the paper.)

A lightweight masked-token model trained on just hours of public data can take any beatbox or finger-tap rhythm and, guided by a short drum-kit sample, generate polished, timbre-matched drum tracks—showcasing an efficient blueprint for controllable audio synthesis.

Q17 (Would you recommend this paper for an award?)

No

Q19 (Potential to generate discourse: The paper will generate discourse at the ISMIR conference or have a large influence/impact on the future of the ISMIR community.)

Agree

Q20 (Overall evaluation: Keep in mind that minor flaws can be corrected, and should not be a reason to reject a paper. Please familiarize yourself with the reviewer guidelines at https://ismir.net/reviewer-guidelines)

Strong accept

Q21 (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

1. Clear technical innovation: Dual-prompt conditioning and the adaptive 2-band dualized rhythm representation cleanly disentangle rhythm from timbre, yielding large gains over 1-band and naïve splits .

2. Convincing empirical evidence with small resources: no significant listener preference between TRIA and MelodyFlow; both strongly preferred over random anchors

3. Thorough ablations and fair crowd-sourcing protocol: Band-count, masking-ratio, and guidance ablations plus IRB-approved listening test with 116 paid evaluators enhance credibility

Weaknesses

1. Narrow evaluation scope – All tests employ ≤ 4 s Western drum prompts, leaving long-form grooves, polyrhythms, and non-Western percussion unvalidated and raising questions about external validity.

2. Prompt-modality mismatch – TRIA receives an audio timbre prompt while MelodyFlow is conditioned on text, introducing a confound that may inflate TRIA’s apparent timbre adherence advantage.

The work delivers a significant, well-substantiated advance: a small, open, energy-efficient model that equals or beats a much larger proprietary system on multiple metrics. Experimental design is solid, writing is clear, and ethical transparency is exemplary. Remaining concerns—limited baselines, short prompts, and unreleased code—are important but fixable.