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.

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

This paper proposes a language modeling framework for drum transcription, one-shot extraction, and reconstruction, based on a new one-shot drum codec.

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 proposes a framework for drum transcription and one-shot extraction, based on a one-shot codec and a language modeling approach. The framework is evaluated on transcription, one-shot extraction, and reconstruction tasks and performs competitively against baselines.

Overall the paper is directly relevant to ISMIR, and the proposed approach is sufficiently novel: there is limited work for joint drum transcription and one-shot extraction, and the proposed language modeling approach has also not been attempted before for these tasks. The paper itself is clear and well written, discusses relevant work, the proposed methodology is technically appropriate and the experiments are well designed. In general this is good work that I recommend for acceptance.

I only have two minor comments:

• Throughout the text there is a mention of "manipulable" or "editable" sequences and representations, however there is no clear description on how can these representations be manipulated and for which purpose. This can be clarified, and perhaps a usecase or example would be useful to have here.

• The structure of section 3 could be improved. The first subsection describes the drum one-shot codec for instance, which according to the pipeline presented in Figure 2, is not the starting point of the process. It would be best if the structure of section 3 would follow more closely the process as depicted in section 3.

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

Summarizing the main comments from all reviews, the paper is generally considered to be interesting and the proposed method is sufficiently novel. There are however specific comments on the evaluation, paper organization, and suggestions on improving the paper's discussion.

Based on all the above, and since a clear accept/reject decision has to be made on each paper, by taking into account all above comments I would recommend acceptance in this case, noting a recommendations on improving the above mentioned points in the camera-ready version.

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

With the right tokenizers and schema, drum transcription and sample extraction can be achieved via language modeling.

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 presents D-Composer, a language model that decomposes drum recordings into manipulable representation containing one-shot samples. Specifically, the model leverages a custom Drum One-shot Codec (DOC) to generate acoustic tokens, alongside additional tokens that represent drum sequences for effective sequence modeling. The evaluation results suggest that the proposed approach performs comparably to baseline systems on tasks such as transcription, one-shot sample reconstruction, and drum sequence reconstruction.

Overall, I found this paper ambitious, and the proposed approach appears novel. Given that Analysis-by-Synthesis (AbS) remains a relatively underexplored paradigm in the context of ADT, the literature review is adequate, and the connections drawn to audio understanding models and multitask systems are interesting. In addition, the demo website provides useful insight into both the audio and transcription quality of the system, while showcasing the model’s ability to capture timbral variations in the extracted one-shot samples.

The main weakness of the paper is the evaluation. While the choice of baseline AbS and tasks-specific systems is reasonable, the experimental setup raises some concerns. The model is trained on a combination of the Lakh and EGMD datasets, while the evaluation includes excerpts from EGMD as well as a “synthetic” set derived from EGMD. This creates the potential for information leakage, which may put competing systems at a disadvantage. Although the authors may have used separate training and testing splits, the comparison is still somewhat unfair given the training procedure employed by the proposed model.

Furthermore, the evaluation of the DOC component is relatively limited. The only reported metric is multi-scale spectral loss (MSL), which does not necessarily correlate well with human perception of audio quality. In fact, based on my experience browsing the examples on the demo website, the perceptual quality of the reconstructed one-shot samples is generally low. If the intended application is professional audio production, a more rigorous assessment of audio quality would be essential and is currently missing from this work.

The writing of the paper is generally clear. However, the reading experience could be improved through better figure placement. For example, Figure 1 appears on the first page but is not referenced until page 4. Similarly, the placement of Figure 2 and 3 could be optimized to better align with the corresponding discussion. In addition, the discussion section feels rushed, and the paper lacks a dedicated conclusion section to effectively summarize the work. While these are relatively minor issues that can be addressed easily, they also suggest that the manuscript would benefit from additional polishing.

In summary, I believe this work is novel and has potential. However, the scope is broad, while the evaluation remains somewhat limited. Taken together with the concerns mentioned above, I believe this paper would benefit from a more comprehensive evaluation as well as a thorough discussion of the results and limitations. At its current stage, my recommendation is a weak reject.

============ Minor comments

Section 1, line 45: “... asdrum…” → as drum?

Section 3, line 187: “... the num of tokens …” → number?

Section 3.1, line 206: Fig. 3 is referred before Fig. 2. It may improve readability to swap their placement or renumber them accordingly.

Section 3.1, line 230-232: out of curiosity, how is the success of the codec defined? Is the primary goal audio quality, compression efficiency, or a balance between the two? From the current description, minimizing the data footprint appears to be the dominant design criterion. Additional discussion or justification would be helpful.

Section 3.2, line 240: I would recommend reconsidering the placement of Figure 1, as it is not referenced until this point.

Section 3.2, line 254: similarly, Figure 2 could be moved closer to its first reference to improve the readability.

Section 3.3, line 274: “... classes inferred by filenames.” → How is the correctness of these assignments verified?

Section 3.3, line 285-286: it would be helpful to report the parameter ranges used for the augmentation methods, as they would provide a sense of the variability.

Section 4.2, line 373: “... random representative one-shot…” → The wording is somewhat unclear. Is the sample randomly selected, chosen to be representative, or both? If both, it would be helpful to clarify the process.

Section 4.3, line 401: “... from unseen Groove MIDI …” → it would be helpful to elaborate on the definition of “unseen”.Does this simply correspond to a train/val/test split, or are additional measures taken to ensure separation by drummer, style, beat type, or other attributes?

Section 5, line 442: “...slightly degraded audio fidelity.” → How is this degradation quantified? Based on the examples provided on the demo page, the degradation appears substantial. Without a subjective listening evaluation or other perceptually motivated assessment, I do not think this claim is sufficiently supported.

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

This study proposes a method for compressing drum one-shot samples into low-dimensional discrete tokens, as well as a method for formulating drum transcription as a problem of recursively estimating discrete tokens representing instrument classes, onset times, and the one-shot samples.

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 and weaknesses

• Strengths: Encoding and transcription methods constructed by combining modern approaches.
• Weaknesses: Paper organization and insufficient discussion (lines 462–473).

Comments

• Why are onset‑time tokens defined up to 10 seconds, even though the drum performance fed into D‑composer is only 4 seconds long?
• Is the proposed method capable of transcribing drum performances of arbitrary length beyond 4 seconds? If so, please describe the specific procedure used to handle longer inputs.
• How are musical notes reconstructed when the predicted token sequence is inconsistent with the expected pattern (i.e., a repetition of one instrument‑type token, one onset‑time token, and thirty‑two one-shot tokens)? For example, how is the reconstruction handled when only thirty one-shot tokens are predicted, or when predicted consecutive onset‑time tokens are in descending order?
• For a fair comparison of compression rates in Table 2, I think that it is necessary to report not only the number of tokens but also the number of bits. For example, your method uses log₂((5⁴)³²) ≈ 297.26 bits to code a single one-shot.

minor errors

• l189: DoC -> DOC
• l45: asdrum -> as drum
• l334: Add a citation for AdamW.
• l529 Codicodec -> CoDiCodec

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

Performing simultaneous transcription and sound event separation via a seq2seq model offers a viable approach to automatic drum transcription

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.

Overall, this paper is excellent and a welcome contribution to the field. What follows are a series of specific points with the aim of improving the existing text.

• Line 2 - I'm not sure I agree with "often" here, but not sure what to suggest instead. Maybe rethink the first sentence?
• Line 12 - change "do not model" to "also fail to" perhaps?
• Line 33 - "one-shots" this might be overloaded in a machine learning context as it can also describe model dynamics e.g. "zero shot performance". "hits" might be an alternative you could use.
• Line 45 - missing space "as_drum"
• Section 2 - related work. Consider also citing "noise2notes" by Yeung et al (ICASSP 2026). Also, your approach of transcription by decomposing into hits and activations is conceptually similar to NMF methods. It would be worth citing Wu and Lerch (ISMIR 2015) as an example of this approach.
• Line 204 - presumably the tradeoff with length 32 is reconstruction quality? Did you try other values? This could be discussed briefly if space allows.
• Line 256 - "we do not expend tokens on silence" - this might be fine for academic evaluation but could pose challenges if this was rolled out as a production system. Maybe reconsider if these tokens are indeed a "waste"?
• Table 2 - this is hard to place in context to get an idea of reconstruction quality without listening to the examples. Perhaps it would be better to include SI-SDR results as these are more commonly used in the source separation literature. It would also be good to include some kind of lower bound method to give a sense of scale e.g. MP3 compression
• Section 4.3 - why did you choose to use 4 second clips as input? Is there scope in this system for stitching the results from consecutive clips together to form a complete transcription for longer excerpts? This is discussed in "Sequence-to-Sequence Piano Transcription with Transformers" by Hawthorne et al. - they use tie tokens to address, but that is for piano which has sustained notes. In any case this should be acknowledged as a shortcoming if a solution is not proposed.
• Table 3 - the results. These are somewhat disappointing compared with ADTOF on the E-GMD and MDB datasets. Listening back to some of the examples on the website it seems like class confusion is an issue. For example the "perc" and "other" classes seem heavily weighted and are likely to be too generic to be useful. I would have liked to see a more principled class system using a restricted 5-class or 8-class set as the model would likely have performed a lot better. The performance on the synthetic data seems positive however.
• Line 446 - it would be nice to give a comment on how you feel the codec sounds here
• References - these would benefit from some more consistent formatting. Please be sure to check them if the work is accepted.

I have attempted to address as many reviewer comments as possible within the revision period given time constraints. I found the feedback very fair and helpful, and plan to address some remaining issues raised by reviewers in follow-up work. While I did not have space to address all questions, some (e.g. windowed inference for recordings >4s) will be clarified by the code release.

Some changes summarized below:

• Clarified and trimmed "manipulable/editable" discussion in line with meta reviewer comments
• Clarified Section 3 in line with meta reviewer comments
• Added discussion of dataset splits and data overlap across methods
• Acknowledged limitations of MSL as reconstruction metric and lack of human evaluations
• Reworked conclusion to clarify results and contributions
• Clarified inference sampling constraints
• Reported codec bitrates in Table 2
• Fixed various typos
• Added citation for Wu & Lerch (NMF drum transcription) and AdamW, corrected CoDiCodec citation
• Added ethics statement & acknowledgements
• Corrected counts in Table 1
• Dropped MusDB mention (previously mentioned but results missing; results now in web supplement; follows same general trend as other datasets, but space too tight to include in Table 3).
• Corrected dataset sizes