Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Pitch Contour Tokenization using VQ-VAE and Its Application on Korean Traditional Music Analysis

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A VQ-VAE trained only to reconstruct fixed-length pitch-contour segments induces a finite codebook of local contour shapes that, without labels, align with expert sigimsae categories and with the two principal pansori modes.

desk verdict VQ-VAE contour tokenization is a real step forward for contour-centric MIR; the mode-alignment evidence needs a null model before it can carry the abstract's weight. read the letter →

arxiv 2608.10979 v1 pith:QRTUVAAC submitted 2026-08-11 cs.SD

Seonguk Ju , Seola Cho , Sooin Chung , Danbinaerin Han , Dasaem Jeong This is my paper · ORCID
classification cs.SD
keywords pitchcontourtokenizationVQ-VAEKoreantraditionalmusicsigimsaepansoridiscreterepresentationlearningornamentanalysistransformation-invariantreconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that the discrete units needed for computational analysis of contour-based music need not be defined in advance by experts: a VQ-VAE trained only to reconstruct fixed-length pitch-contour segments can induce a finite codebook of local contour shapes from unlabeled audio. The authors argue that these tokens are stable across segmentation positions and small variations in timing and pitch range because the reconstruction loss is minimized over candidate temporal and pitch-domain transformations. They demonstrate that the learned tokens carry label-discriminative information for expert-defined sigimsae ornament categories without any supervision, and that in pansori individual tokens align with the two principal modes, Gyemyeonjo and Ujo. If correct, the approach supplies corpus-level analysis with a unit for contour-centric traditions that plays the role the note plays for score-based music.

What carries the argument

The central mechanism is the transformation-minimized reconstruction loss, which evaluates reconstruction quality after searching over temporal scale, temporal offset, pitch-range scaling, and pitch shift and backpropagates only through the best alignment. This is what lets the codebook group contours by shape rather than by phase, timing, or range. The other load-bearing design is receptive-field separation with median subtraction: the input is pre-segmented into non-overlapping 128-frame windows so each token corresponds to exactly one fixed-length contour segment, and subtracting the segment median removes absolute pitch so tokens encode local movements such as oscillations and slides.

What would settle it

A synthetic control would settle the shape-invariance claim: generate contour pairs that share the same melodic shape but differ only in absolute pitch level, vibrato width, or segmentation phase, and check whether token assignment is preserved; if tokens flip under pitch shift or range scaling, the transformation-minimized loss has not removed non-shape factors. An equivalent real-data check would tokenize the same pansori phrase transposed by several semitones and compare codebook usage.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a discrete vocabulary of local pitch-contour patterns can be induced directly from unlabeled F0 contours. Median normalization strips absolute pitch from each receptive-field segment, the transformation-minimized reconstruction loss absorbs segmentation phase and small timing and pitch-range differences, and the resulting 256-entry codebook assigns the same token to shape-equivalent contours. In the authors' evaluation, a single token selected per sigimsae class visually matches the canonical realization of that class, and in pansori one token separates Gyemyeonjo from Ujo by vibrato width on an otherwise identical descending motion, a distinction that note-level transcription would discard.

Load-bearing premise

The load-bearing premise is that median normalization plus the pitch-shift and pitch-range reconstruction transforms remove every non-shape factor, so a token that appears predominantly in Ujo must be responding to contour shape (here vibrato width) and not to residual pitch level, range, or F0-extraction artifacts.

Editorial extensions

If this is right

  • Dense token sequences can be computed over entire voiced contours, making transition statistics and n-gram analyses available for contour-centric traditions in the way they have long been for chord and note sequences.
  • Without any ornament labels, the most representative token for each sigimsae class visually matches the canonical realization of that class, and the full transformation-minimized loss improves per-token classification over the autoencoder baseline.
  • In pansori, token 54 separates Gyemyeonjo from Ujo by the width of vibrato on an otherwise identical descending motion, preserving an ornamental distinction that note-level transcription discards.
  • The same codebook, queried against other label sets such as schools or performers, offers a path toward corpus-level study of ornamental practice without predefined ornament categories.
  • Because the method requires no labels and no note-level transcription, it applies in principle to other contour-centric traditions, including Indian art music and pop vocal singing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct control the paper does not run: synthesizing contour pairs identical in shape but differing only in vibrato width or absolute pitch level would test whether token 54's mode separation is caused by contour shape rather than by a residual pitch or range artifact.
  • Because the tokenizer discards segments with extensive unvoiced frames, corpus-level statistics built on it will systematically under-represent phrases with weak or missing voicing, a coverage gap the authors acknowledge.
  • The paper demonstrates per-token mode association but not that token transition statistics differ across modes; testing n-gram or transition matrices over the same codebook across Gyemyeonjo and Ujo would show whether the units support the corpus-level analyses the paper motivates.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes an unsupervised VQ-VAE tokenizer for fixed-length pitch-contour segments. To promote segmentation and pitch-range robustness, the reconstruction loss is minimized over candidate temporal offsets, resampling scales, pitch shifts, and pitch-range scalings. The method is evaluated on Korean traditional music through segmentation-shift consistency (Section 5), a per-token MAP probe for sigimsae categories (Section 6), and a qualitative mode analysis in pansori (Section 7). The paper claims that the learned codebook is stable, that its tokens recover label information without supervision, and that individual tokens can separate Gyemyeonjo and Ujo.

Significance. If substantiated, the proposed tokenization would be a valuable tool for corpus-level analysis of traditions organized around continuous pitch movement, because the discrete vocabulary is induced from unlabeled audio rather than predefined by experts. The paper has clear strengths: the transformation-minimized loss is a well-motivated architectural contribution, the code and demo are publicly available, and the deliberately weak probing classifier in Section 6 makes the sigimsae result a conservative test of label information. The segmentation-shift consistency results in Table 1 are internally coherent. However, the two strongest interpretive claims—that tokens carry mode information and that the learned invariances remove pitch-range confounds—currently rest on evidence without null-model controls or statistical support.

major comments (3)
  1. [Section 7, Figure 3] The mode alignment is established by selecting, for each mode, the token with the highest P(mode|token) from a 256-entry codebook. This is a post hoc maximum: under the null hypothesis that tokens carry no mode information, the maximum of 256 estimated probabilities is expected to be far above the base rate. No permutation test, null-model comparison, or multiple-comparison correction is reported. The claim that 'individual tokens align with the two principal modes' therefore lacks quantitative support. Please report the full distribution of P(mode|token) over the codebook, the counts behind the token 54 example, and the significance of the selected token relative to a null model in which segment labels are permuted.
  2. [Section 7; Section 3.2, Eq. (4)] The paper states that because tokens are median-normalized, 'any mode-specific token usage must arise from contour shape itself.' Median subtraction removes absolute pitch offset, but not pitch range or amplitude, and Eq. (4) explicitly includes pitch-range scaling with A={0.85,1.0,1.15}, making pitch-range invariance an intended property rather than a guaranteed one. Table 1 shows that adding the pitch-domain transformations degrades segmentation-shift consistency (KLD 0.652 vs. 0.531 for temporal alignment alone), so residual sensitivity to pitch range is plausible. The wide-vibrato reading of Token 54 should be checked with a control: rescaled pitch-range versions of the same contours should be fed to the tokenizer and the token assignments compared, and Token 54's assigned contours should be compared to range-matched contours from both modes. Without this, the mode separation may be a residual range effect rather than vibrato shape.
  3. [Section 6.3, Table 2] The reported improvements of the full transformation loss over the autoencoder baseline on the vocal-only setting are small (F1 0.331 vs. 0.324; mAP 0.324 vs. 0.290), and all results appear to come from a single training run. No error bars, repeated seeds, or significance tests are provided. Given that the sigimsae probing experiment is the main quantitative evidence that the learned tokens 'recover information about expert-defined sigimsae categories,' the reader cannot currently distinguish a genuine improvement from training noise. Please add multiple seeds with confidence intervals or a statistical test over runs.
minor comments (5)
  1. [Section 6.2, Figure 2] The quantities labeled 'P=0.02', 'P=0.04', etc. in Figure 2 are conditional token probabilities P(token|label), not p-values; please relabel them to avoid statistical misinterpretation.
  2. [Section 4.1] The relationship among the 280-hour in-house pansori dataset, the 34.0/10.0-hour AI-Hub sigimsae set, and the 26.7/8.7-hour mode-annotated subset should be stated explicitly, in particular whether the mode subset is disjoint from the training segments used in Sections 5 and 6.
  3. [Section 3.2, Eq. (4)] The text says that only the transformation attaining the minimum error is used for backpropagation, but it would be helpful to state explicitly that the selected s, tau, a, and b are fixed when computing gradients, to rule out any confusion about differentiating through a minimum.
  4. [Section 7] The phrase 'Tokens 171' should be 'Token 171' for consistency with the surrounding text and with other token references.
  5. [Throughout] The model name appears as 'VQ-V AE' with a space in several places; this typography should be corrected to 'VQ-VAE'.

Circularity Check

1 steps flagged · score 5.0 of 10

Pansori mode-alignment evidence is a post hoc maximum over the codebook: selecting the token with highest P(mode|token) makes 'occurs predominantly in Ujo' true by construction, while the core VQ-VAE and sigimsae probe remain non-circular.

  1. self definitional [Section 7, 'Token Analysis in Pansori' (method and Token 54 discussion)]
    "We therefore count token occurrences within mode-labeled segments and identify the tokens with the highest P(mode|token) for each mode, treating these as candidate ornamental patterns characteristic of each mode. ... Token 54 is particularly informative. ... yet this specific token occurs predominantly in Ujo."

    Token 54 is selected precisely because it maximizes P(Ujo|token) over the 256-entry codebook, so the statement that this specific token 'occurs predominantly in Ujo' is true by construction, not by an independent test. The paper then presents this selection as evidence that 'the tokenizer, trained without any mode supervision, separates the two modes along this ornamental dimension.' No null model, permutation test, or multiple-comparison correction is reported, so the apparent alignment is guaranteed by the selection criterion. The only non-tautological residue is the post hoc visual/score interpretation of the selected contours, which is not statistically evaluated.

full rationale

The core tokenization pipeline is self-contained: the VQ-VAE is trained with a reconstruction objective on unlabeled F0 contours (Section 3), and the sigimsae probe (Section 6) is a legitimate held-out evaluation — P(token|label) is estimated on the training split and applied to test segments, so the F1/mAP results are not circular. The segmentation-shift consistency test (Section 5) is likewise an internal evaluation of the model's own robustness, not a relabeling of inputs. The only material circularity is in the pansori mode analysis (Section 7): the tokens presented as aligning with Gyemyeonjo and Ujo are chosen by maximizing P(mode|token), so the statement that the selected token 'occurs predominantly in Ujo' is true by construction and is not an independent confirmation of mode separation. Citation [24] is a self-citation (anonymous submitted paper) used as the source of mode annotations, but since the annotations are expert labels rather than token outputs, this is not in itself circular. Thus, taking the weak claim — the learned codebook carries enough shape information to make held-out sigimsae classification possible — the paper is not circular; taking the strong claim — that individual tokens align with the two principal modes — the evidence reduces to the selection criterion.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on preprocessing and architectural assumptions: F0 extraction quality, the existence of local contour shapes that fit in a 128-frame window, the reliability of expert labels, and the paper's explicit decision to discard segments with unvoiced frames. The transformation candidate sets and MAP temperature are hand-chosen or fitted. No new physical entities are introduced; codebook tokens are learned representational units.

free parameters (4)
  • Transformation candidate sets S, A, B = S={1,1.08,1.2}; A={0.85,1.0,1.15}; B={-0.1,0.0,0.1}
    Chosen empirically in Section 4.2. These sets determine how much temporal offset, tempo, pitch shift, and pitch-range variation the reconstruction loss tolerates, so they directly shape what the codebook treats as the same contour pattern.
  • MAP classifier temperature tau = Selected on validation set in [0,1]
    Section 6.2; the temperature balances the label prior against the token likelihood and is tuned on validation labels, affecting the reported F1 and mAP numbers.
  • Receptive field size and codebook size = 128 frames (1.28 s) and 256 entries
    Section 4.2; the 1.28 s window must contain complete ornamental gestures for the token-label probing to be meaningful, and the codebook size sets the granularity of the vocabulary.
  • Unvoiced-segment filtering and CREPE confidence threshold = Not specified numerically
    Section 4.1.1; segments with extensive unvoiced frames are discarded and gaps up to two frames are interpolated. This systematically removes parts of recordings and may bias the learned vocabulary, and the paper itself lists this as the main limitation.
assumptions (6)
  • standard math VQ-VAE training objective with straight-through estimator and nearest-neighbor lookup is a valid way to learn discrete codes
    Invoked in Section 2.2, Eq. (1); this is the standard VQ-VAE formulation from van den Oord et al.
  • domain assumption CREPE F0 estimates at 10 ms resolution, after confidence filtering, are accurate enough to represent sigimsae ornaments
    Section 4.1.1 uses pretrained CREPE as the sole pitch source; errors or octave jumps would enter the tokens directly.
  • domain assumption Sigimsae ornaments are local contour shapes that fit within a 128-frame receptive field and remain separable after median normalization
    Sections 3.1 and 6.2 crop each annotated segment to the receptive field and encode a single token; if ornaments span longer contexts, this probing setup cannot capture them.
  • domain assumption Expert annotations in the NIA AI-Hub sigimsae dataset and the anonymous mode-annotated subset [24] are reliable
    Sections 6.1 and 7 treat the labels as ground truth; [24] is an anonymous submitted paper, so the annotation protocol cannot be inspected.
  • ad hoc to paper Discarding segments with extensive unvoiced frames does not bias the analysis
    Section 4.1.1 and the conclusion state the 1D-CNN tokenizer cannot model unvoiced regions and that training/evaluation were restricted to dense voiced segments; this is a stated limitation of the method.
  • domain assumption Mode-specific token usage arises from contour shape because tokens are median-normalized
    Section 7 argues that since tokens carry no absolute pitch, mode-specific usage must come from shape; this ignores possible residual correlations with pitch range, phase, or F0 artifacts, and is not tested with a control.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pitch Contour Tokenization using VQ-VAE and Its Application on Korean Traditional Music Analysis." pith.science (2026). https://pith.science/paper/QRTUVAAC

@misc{pith2026260810979,
  author       = {Pith},
  title        = {Pith review of: Pitch Contour Tokenization using VQ-VAE and Its Application on Korean Traditional Music Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QRTUVAAC}},
  note         = {Machine review of arXiv:2608.10979}
}
read the original abstract

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.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

  1. [1]

    Pitch Contour Tokenization using VQ-VAE and Its Application on Korean Traditional Music Analysis

    INTRODUCTION Representing music as sequences of discrete units has been one of the central abstractions underlying computational musicology and symbolic music information retrieval. Its analytical value lies in the fact that transitions between mu- sical events can be counted and modeled statistically. For example, corpus-based chord progression analysis ...

    work page Pith review arXiv 2026

  2. [2]

    Ranjani et al

    RELATED WORK 2.1 Pitch-Contour Abstractions for Ornamental Music A notable precedent is found in Indian art music, where quantized pitch contours have served as an intermediate representation capturing ornamental detail beyond note- 1 Codebase:https://github.com/SeongUkJu/pitch-con tour-tokenizer, Demo:https://seongukju.github.io/ pitch-contour-tokenizer-...

    work page

  3. [3]

    VQ-V AE ON PITCH CONTOUR The overall architecture is based on VQ-V AE, using a stack of 1D convolutional layers, with two key de- sign choices to enable interpretable, segmentation-robust (position-invariant) tokenization of pitch contours, as illus- trated in Figure 1. 3.1 Receptive Field Separation To ensure a 1:1 correspondence between each discrete to...

    work page

  4. [4]

    Although the proposed method is applicable to any pitch contour-based musical tradition, such as Indian art music or pop vocal singing, we focus on Korean traditional music

    EXPERIMENTAL SETUP 4.1 Dataset In this study, we apply VQ-V AE to pitch contours of Ko- rean traditional music. Although the proposed method is applicable to any pitch contour-based musical tradition, such as Indian art music or pop vocal singing, we focus on Korean traditional music. This is because continuous pitch movement is not merely an expressive e...

    work page

  5. [5]

    + temporal align

    SEGMENTATION SHIFT CONSISTENCY We introduced the transformation-minimized reconstruc- tion loss to reduce the sensitivity of token assignment to temporal and pitch-domain variations. Here, we isolate one of its intended effects: robustness to segmentation position. Specifically, we test whether the same long pitch contour yields similar token sequences wh...

    work page

  6. [6]

    We test this with a minimal probing setup, a per- token MAP classifier — that isolates how much label- discriminative information is carried by individual tokens

    SIGIMSAE CLASSIFICATION If the learned tokens are to serve as analytical units, their patterns should align with expert-defined ornament cate- gories. We test this with a minimal probing setup, a per- token MAP classifier — that isolates how much label- discriminative information is carried by individual tokens. 6.1 Sigimsae Labeled Dataset We employed th...

    work page

  7. [7]

    Dataset:Pansorihas two representative modes:Gye- myeonjoandUjo, each characterized by distinct melodic and ornamental patterns

    TOKEN ANALYSIS IN PANSORI As a case study, we explore whether the learned tokenizer can serve as a tool for data-driven musicological analysis by examining how individual tokens relate to the two prin- cipal modes ofpansori. Dataset:Pansorihas two representative modes:Gye- myeonjoandUjo, each characterized by distinct melodic and ornamental patterns. We u...

    work page

  8. [8]

    CONCLUSION AND FUTURE WORK We proposed a VQ-V AE-based tokenizer that learns a dis- crete vocabulary of local pitch-contour patterns without or- nament labels, together with a transformation-minimized reconstruction loss that makes the learned tokens robust to small temporal and pitch-domain variations. Experiments on Korean traditional music show improve...

    work page

Show all 34 references
  1. [9]

    ACKNOWLEDGEMENT This work was supported by the National Research Foun- dation of Korea (NRF) grant funded by the Korea govern- ment (MSIT) (RS-2025-00560548)

    2025

  2. [10]

    Statistical characteristics of tonal harmony: A corpus study of beethoven’s string quartets,

    F. C. Moss, M. Neuwirth, D. Harasim, and M. Rohrmeier, “Statistical characteristics of tonal harmony: A corpus study of beethoven’s string quartets,”PLoS One, vol. 14, no. 6, p. e0217242, 2019

    2019

  3. [11]

    Information flow and repetition in mu- sic,

    D. Temperley, “Information flow and repetition in mu- sic,”Journal of Music Theory, vol. 58, no. 2, pp. 155– 178, 2014

    2014

  4. [12]

    In search of the horowitz factor,

    G. Widmer, S. Dixon, W. Goebl, E. Pampalk, and A. Tobudic, “In search of the horowitz factor,”AI Mag- azine, vol. 24, no. 3, pp. 111–111, 2003

    2003

  5. [13]

    Yourmt3+: Multi-instrument music transcription with enhanced transformer architectures and cross-dataset stem augmentation,

    S. Chang, E. Benetos, H. Kirchhoff, and S. Dixon, “Yourmt3+: Multi-instrument music transcription with enhanced transformer architectures and cross-dataset stem augmentation,” in2024 IEEE 34th International Workshop on Machine Learning for Signal Processing (MLSP), 2024, pp. 1–6

    2024

  6. [14]

    Svara-forms in carnatic music: Contextual influences on the perfor- mance of svara,

    T. Nuttall, X. Serra, and L. Pearson, “Svara-forms in carnatic music: Contextual influences on the perfor- mance of svara,” inIEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2025 - Workshops, Hyderabad, India, April 6-11, 2025, 2025, pp. 1–5

    2025

  7. [15]

    Beyond notation: A digital platform for transcribing and analyzing oral melodic traditions,

    J. Myers and D. Neuman, “Beyond notation: A digital platform for transcribing and analyzing oral melodic traditions,” inProceedings of the 26th International Society for Music Information Retrieval Conference, 2025, pp. 421–429

    2025

  8. [16]

    Quantized melodic contours in indian art music per- ception: Application to transcription,

    H. G. Ranjani, D. Paramashivan, and T. V . Sreenivas, “Quantized melodic contours in indian art music per- ception: Application to transcription,” inProceedings of the 18th International Society for Music Information Retrieval Conference, ISMIR 2017, Suzhou, China, October 23-...

    2017

  9. [17]

    A compact pitch and time rep- resentation for melodic contours in indian art music,

    H. G. Ranjani, A. Srinivasamurthy, D. Paramashivan, and T. V . Sreenivas, “A compact pitch and time rep- resentation for melodic contours in indian art music,” The Journal of the Acoustical Society of America, vol. 145, no. 1, pp. 597–603, 01 2019

    2019

  10. [18]

    Hierarchical generative modeling of melodic vocal contours in hindustani classical music,

    N. N. Shikarpur, K. M. Dendukuri, Y . Wu, A. Caillon, and C. A. Huang, “Hierarchical generative modeling of melodic vocal contours in hindustani classical music,” inProceedings of the 25th International Society for Music Information Retrieval Conference, ISMIR 2024, San Franci...

    2024

  11. [19]

    Mining melodic patterns in large audio collections of indian art music,

    S. Gulati, J. Serrà, V . Ishwar, and X. Serra, “Mining melodic patterns in large audio collections of indian art music,” in2014 Tenth International Conference on Signal-Image Technology and Internet-Based Systems, 2014, pp. 264–271

    2014

  12. [20]

    The matrix profile for motif discovery in audio - an example application in carnatic music,

    T. Nuttall, G. Plaja-Roglans, L. Pearson, and X. Serra, “The matrix profile for motif discovery in audio - an example application in carnatic music,” inMusic in the AI Era - 15th International Symposium, CMMR 2021, Virtual Event, November 15-19, 2021, Revised Selected Papers, ...

    2021

  13. [21]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, 2017, pp. 6306–6315

    2017

  14. [22]

    Symbolic music loop generation with neural discrete representations,

    S. Han, H. Ihm, M. Lee, and W. Lim, “Symbolic music loop generation with neural discrete representations,” inProceedings of the 23rd International Society for Music Information Retrieval Conference, ISMIR 2022, Bengaluru, India, December 4-8, 2022, 2022, pp. 403– 410

    2022

  15. [23]

    Jukebox: A generative model for mu- sic,

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, “Jukebox: A generative model for mu- sic,”CoRR, vol. abs/2005.00341, 2020

    2005 arXiv

  16. [24]

    Soundstream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “Soundstream: An end-to-end neural audio codec,”IEEE ACM Trans. Audio Speech Lang. Process., vol. 30, pp. 495–507, 2022

    2022

  17. [25]

    High fidelity neural audio compression,

    A. Défossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,”Trans. Mach. Learn. Res., vol. 2023, 2023

    2023

  18. [26]

    Unsupervised speech representation learning using wavenet autoencoders,

    J. Chorowski, R. J. Weiss, S. Bengio, and A. van den Oord, “Unsupervised speech representation learning using wavenet autoencoders,”IEEE ACM Trans. Audio Speech Lang. Process., vol. 27, no. 12, pp. 2041–2053, 2019

    2019

  19. [27]

    Hybrid trans- formers for music source separation,

    S. Rouard, F. Massa, and A. Défossez, “Hybrid trans- formers for music source separation,” inIEEE Inter- national Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023, 2023, pp. 1–5

    2023

  20. [28]

    Crepe: A convolutional representation for pitch estimation,

    J. W. Kim, J. Salamon, P. Li, and J. P. Bello, “Crepe: A convolutional representation for pitch estimation,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2018, Calgary, AB, Canada, April 15-20, 2018, 2018, pp. 161–165

    2018

  21. [29]

    Svara-forms and coarticulation in carnatic music: an investigation using deep clustering,

    T. Nuttall, X. Serra, and L. Pearson, “Svara-forms and coarticulation in carnatic music: an investigation using deep clustering,” inProceedings of the 11th Interna- tional Conference on Digital Libraries for Musicology, DLfM 2024, Stellenbosch, South Africa, 27 June 2024, 2024...

    2024

  22. [30]

    UMAP: uniform manifold approximation and projection for dimension reduc- tion,

    L. McInnes and J. Healy, “UMAP: uniform manifold approximation and projection for dimension reduc- tion,”CoRR, vol. abs/1802.03426, 2018

    2018 arXiv

  23. [31]

    Ko- rean traditional music score and audio dataset,

    The Open AI Dataset Project (AI-Hub, S. Korea), “Ko- rean traditional music score and audio dataset,” AI- Hub, 2022, dataset No. 71470. Available at https://ai hub.or.kr/aihubdata/data/view.do?dataSetSn=71470

    2022

  24. [32]

    Recognizing orna- ments in vocal indian art music with active annotation,

    S. Kumar, P. Singh, and V . Arora, “Recognizing orna- ments in vocal indian art music with active annotation,” CoRR, vol. abs/2505.04419, 2025

    2025 arXiv

  25. [33]

    Frame-level pansori mode classification with complementary audio representations,

    Anonymous, “Frame-level pansori mode classification with complementary audio representations,” 2026, sub- mitted to ISMIR 2026

    2026

  26. [34]

    A study on the Shigimsae of Pansori Kye- myunjo,

    J. seon Jo, “A study on the Shigimsae of Pansori Kye- myunjo,” Master’s thesis, Hanyang University, Seoul, Republic of Korea, 2000, original title in Korean:판소 리계면조시김새에관한고찰[조주선]

    2000

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.