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.

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

This paper proposes a VQ-VAE framework applied to two kinds of Korean Music to create pitch contour tokens for use in the later downstream segmentation task.

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 was a joy to read. While I do not work on Korean music, I felt that the authors wrote so accessibly as to invite me into understanding their work. I appreciated that on the bottom of the first page, they plainly stated the assumption of "note-units" for many AMT systems because these systems are largely based on Western music. For me this was a critical framing that set the stage both for their approach in this work and for the potential impact this work could have on the MIR community more broadly.

Throughout this paper, I felt that the authors did an excellent job motivating their approach and their assumptions. At the top of Section 3.2, I was initially taken aback by their choice to have non-overlapping windows, but then as if the authors could hear my hesitation, the very next clause addressed my concern. This happened again in Section 5.3 when I was wondering why adding the pitch offset transformations would cause the system to perform worse (as shown in Table 1). Again, the authors offered a clear and accessible explanation.

I found this paper to be clever in many ways, from their application of VAEs and their choice of musical data to their loss function and their choices of baseline comparisons. The loss function reminded me of previous MIR work (specifically work by Goto et al.) that sought to find the best candidate for comparison in a pitch invariant manner. I also thought it was brave for the authors to include ED-TCN as a comparison to their methods because ED-TCN outperfomed their proposed methods. But again, the authors were quick to note that ED-TCN is a supervised method, which is in sharp contrast to their proposed unsupervised method. The authors cleverly note that they included this comparison to demonstrate the gap between their proposed method, which improves on the performance of other unsupervised methods, and supervised methods.

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

In general the reviewers were impressed by this work. We found the paper to be well motivated and relevant to ISMIR broadly. We encourage the authors to read our comments carefully, but I will highlight a few items in particular:

• Reviewer 2 raised a few issues regarding unjustified choices regarding various parameters.
• All the reviewers noted the performance for the vanilla VQ-VAE
• Reviewer 3 noted that both the AI usage section and the Ethics section are missing.

Again, in general we found this work to be compelling and appropriate for ISMIR.

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

Pitch contours in traditional music can be automatically learned and improve analysis of said music.

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 was a delight to read, presented a compelling case for the proposed method (human expert labelling being used in other works), and was written to be easy to understand as someone with little knowledge of Korean traditional music.

I am least sure regarding the evaluation, and whether the baseline classification method is comparable, however it seems sensible to me as an base-level evaluation.

I hope to see further analysis performed using this method, and the tool made available.

A small useful addition to this paper would be linking to an online source visualising or containing audio examples of the learned features.

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

A VQ-VAE with transformation-minimized reconstruction loss learns an interpretable, segmentation-robust vocabulary of pitch-contour patterns from unlabeled pansori audio, recovering meaningful ornamental categories without supervision.

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 using VQ-VAE to learn a discrete vocabulary of local pitch-contour patterns from unlabeled audio, targeting musical traditions where continuous pitch movement resists note-level segmentation. The main technical contribution is a transformation-minimized reconstruction loss that makes the learned tokens robust to shifts in segmentation position and to small pitch-domain variations. The method is evaluated on pansori (Korean traditional music) through three experiments: segmentation-shift consistency, sigimsae (ornament) classification using a simple token-lookup classifier, and qualitative analysis of tokens associated with the two principal pansori modes (Gyemyeonjo and Ujo).

The problem formulation is the paper's strongest aspect. The authors make a convincing case that many musical traditions are built around continuous pitch movement where the "discrete units" that analysis would build on simply don't exist in advance. Prior approaches either discretize pitch values onto predefined grids (Ranjani et al., Shikarpur et al.) or rely on manually specified trajectories (IDTAP). Learning a codebook of contour shapes directly from data is a natural and well-motivated alternative, and I was not able to find prior work that does this in the way proposed here.

The transformation-minimized reconstruction loss is a clean technical idea. Standard frame-level MSE penalizes segmentation-induced misalignments even when the underlying pattern is the same, which would push the model toward encoding positional information rather than shape. Minimizing over candidate temporal and pitch-domain transformations before computing the loss is an elegant solution. The formulation in Equations 2 through 4 is mathematically consistent, and the hard-min selection for backpropagation is straightforward to implement.

I appreciate the honesty of the experimental reporting. Vanilla VQ-VAE does not outperform the autoencoder baseline on sigimsae classification (Table 2), and the paper says so plainly. It is only with the full transformation-minimized loss that the model pulls ahead. The interplay between Tables 1 and 2 is also interesting: temporal alignment alone is best for segmentation consistency, while all transformations are needed for classification. The paper explains this coherently, noting that pitch-domain transformations are not designed for segmentation robustness but help the codebook capture contour shape. The pansori mode analysis in Section 7 provides good qualitative evidence that the learned tokens carry musicologically meaningful information. The finding about Token 54, which captures a melodic move shared by both modes but distinguished by the vibrato pattern on the upper note, is particularly compelling as a case where contour-level analysis reveals structure that note-level transcription would miss.

However, there are several issues I want to raise. The paper fixes two critical hyperparameters, codebook size (K=256) and receptive field (L=128 frames), without any ablation or justification. These choices fundamentally determine what the model can learn: the codebook size controls vocabulary expressiveness, and the receptive field determines the temporal scale of captured patterns. For a methods paper proposing a general framework, I would expect at least a basic sensitivity analysis showing how results change at, say, K=64 or K=512, or at receptive fields of 64 or 256 frames. The transformation parameters (resampling rates, pitch-range scales, pitch offsets) are also set empirically without discussion of sensitivity. The commitment coefficient beta=2.0 is at the extreme upper end of values used in the VQ-VAE literature (van den Oord et al. used 0.25 in their experiments) and no reasoning is given.

Also, only one unsupervised baseline is compared (autoencoder with UMAP and K-Means, adapted from Nuttall et al.). It would be helpful to see simpler alternatives like K-Means directly on raw windowed contours, or other vector quantization variants, to better isolate where the gains come from. Is it the VQ-VAE architecture, the transformation loss, or simply having any learned codebook? The current experiments don't fully disentangle this.

Finally, the gap to the supervised ED-TCN on sigimsae classification is large. The paper describes the best unsupervised result as recovering "roughly two-thirds" of the supervised F1, but the actual ratio (0.342/0.541 = 63.2%) is closer to three-fifths. The hedging word "roughly" makes this acceptable, but it is worth noting that the practical utility for automatic ornament classification remains limited at this performance level. The inability to handle unvoiced regions also restricts the model to carefully filtered segments, which limits applicability to full recordings.

Despite these concerns, I think the core contribution is sound and the paper is well above average in writing quality and intellectual honesty. The idea of contour tokenization for non-Western music analysis is timely and relevant to ISMIR, and the transformation-minimized loss is a genuine methodological contribution.

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

a model that learns recurring shapes on its own, without ornament types, and matches expert sigimsae categories and reveal differences otherwise missed by 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.

the paper proposes an unsupervised pitch contour tokenizer. strengths well motivated and ismir relevant : the transformation minimized loss is a well argues idea. the results are convincing, token-54 showing structure otherwise lost by transcription. clear positioning against grid quantization and honest about its limits. weaknesses: the quantitative gains are small. no variance. autoencoder baseline only beats by 1.5 to 2.7 f1 pts. the vanilla vq-vae is worse - would want to see mean and standard deviation over seeds reported with a significance test. there are two evaluations which favor different configurations. maybe reporting them both consistently and presenting the tradeoff as a finding would be better mode analysis is one qualitative case wasnt sure about the transofmration sets - hard to follow - especially resampling at 50, 54, 60 hz relative to 100hz. didnt include the ai section or ethics section.

We thank the reviewers for their constructive and encouraging feedback. The following changes were made after the review phase:

1. Code release and audio demo. As suggested by Reviewer 1, we now publicly release our implementation and provide an online demo page containing audio examples and visualizations of the learned tokens, so that readers can directly inspect and listen to the discovered contour patterns.

2. Clarified the temporal-resampling description. Reviewer 3 noted that the description of the resampling transformation was hard to follow. We rewrote this passage to express the resampling directly in terms of the input length L: the decoder output (2L = 256 frames) is resampled to lengths of approximately 1, 1.08, and 1.2 times L (128, 139, and 154 frames). This removes the unintuitive frequency notation and makes the temporal-scale candidates explicit.

3. AI Usage Statement. Under the ISMIR 2026 AI Usage Policy, the use of programming aids and grammar-correction and language editing tools for editing purposes does not require declaration. Accordingly, we did not add a separate AI Usage Statement. No AI tools were used beyond these purposes, and the authors take full responsibility for the correctness and originality of all content in the paper.

4. Results updated to match the released checkpoints. While preparing the code release, we additionally fixed the random seed and enabled deterministic cuDNN operations, which were uncontrolled in the submitted version, and updated Tables 1–2 so that every reported number matches a released checkpoint. The minor differences from the submitted numbers arise mainly from non-deterministic, multi-threaded k-means initialization of the VQ codebook (the autoencoder baseline, which applies k-means only after training, was deterministic). We therefore release the trained weights so that all scores can be reproduced directly from the checkpoints. All trends and conclusions are unchanged.