Pith. sign in

REVIEW 3 major objections 4 minor 40 references

Multi-Class-Token Transformer for Multitask Self-supervised Music Information Retrieval

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single 5.3-million-parameter transformer with two class tokens—one contrastive, one equivariant—beats both single-objective baselines and a 95-million-parameter masked-language model on five of six music tasks.

desk verdict Useful multi-token SSL paper for MIR, but the central mechanism is confounded with input fusion and the abstract overclaims on key and chord. read the letter →

arxiv 2507.12996 v1 pith:IYAX33LU submitted 2025-07-17 cs.SD

Yuexuan Kong , Vincent Lostanlen , Romain Hennequin , Mathieu Lagrange , Gabriel Meseguer-Brocal This is my paper · ORCID
classification cs.SD
keywords self-supervisedlearningmusicinformationretrievalcontrastiveequivariantclasstokenvisiontransformermultitaskcircleoffifths
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

Self-supervised learning for music has a known trade-off: contrastive objectives (pulling same-track segments together, pushing other tracks apart) are good at tagging but weak at structured predictions like key, while equivariant objectives (regressing a known transformation such as a pitch shift) are strong on their target task but do not generalize. The paper proposes a single small transformer that runs both kinds of pretext at once, giving each its own class token inside the same backbone. On six downstream tasks, averaging the two tokens yields a 5.3-million-parameter model that beats both single-objective baselines and a 95-million-parameter masked-language-model baseline on five tasks under identical linear probing. The point of the work is to show that multitask self-supervision with multiple class tokens is a parameter-free way to combine complementary invariances—timbre invariance from contrastive learning and pitch-class structure from equivariant learning—into one representation.

What carries the argument

The central mechanism is the multi-class-token multitask (MT2) ViT-1D: a single Vision Transformer with 1-D spectrogram patches whose input is a sum of mel and CQT patch embeddings plus two learnable class tokens, one optimized by the contrastive NT-Xent loss and the other by the CPSD equivariant loss. The CQT pathway adds a residual connection from the average CQT frame to the equivariant token before octave-pooling to a 12-dimensional softmax chroma vector, which is what the CPSD distance over the circle of fifths actually measures. The contrastive pathway linearly projects its token to 512 dimensions before NT-Xent. Sharing one backbone while keeping the tokens separate is what lets the model combine timbre-invariant and pitch-structured representations without extra parameters; the token average at probing time is the fusion mechanism.

What would settle it

On a held-out test set of atonal or percussive tracks, record the entropy of the model's 12-dimensional pitch-class output alongside the equivariant token's key-estimation accuracy: if high-entropy outputs still predict keys correctly, the one-hot assumption is not load-bearing, whereas if accuracy collapses, the equivariant guarantee is conditional on tonal material.

Watch

Extended reading notes

Core claim

The paper's discovery is that a transformer does not need to choose between contrastive and equivariant pretraining: it can run both in the same forward pass by giving each pretext a separate class token. The MT2 ViT-1D feeds the transformer with the sum of a mel-spectrogram patch sequence and a constant-Q transform patch sequence, prepended by two learnable tokens. The contrastive token is projected to a 512-dimensional vector and trained with NT-Xent against same-track positives and cross-track negatives; the equivariant token is projected through octave pooling and softmax to a 12-dimensional pitch-class vector and trained with the cross-power spectral density (CPSD) loss, including a pitch-shifted third segment so the circle-of-fifths shift is supervised. The two tokens specialize cleanly—one forms a circle-of-fifths ring across instruments, the other clusters by instrument and is pitch-invariant—and, averaged or taken individually, they outperform the single-token monotask models and MERT's last layer on all downstream tasks except beat tracking, with 18x fewer parameters.

Load-bearing premise

The load-bearing premise is that the equivariant loss can reduce real audio to a single confident pitch class, since the circle-of-fifths guarantee holds only for near-one-hot pitch-class vectors; diffuse, uncertain pitch-class distributions in non-tonal music would weaken the key and pitch gains.

Editorial extensions

If this is right

  • A frozen 5.3M-parameter representation, probed with one linear layer, reaches or beats a 95M-parameter masked-language model on music tagging, instrument recognition, key estimation, pitch estimation, and chord estimation.
  • Because the equivariant token improves the contrastive token's tagging and instrument scores relative to the monotask contrastive model, joint training transfers harmonic structure into the contrastive stream without task-specific fine-tuning.
  • Averaging the two class tokens is a parameter-free fusion that improves key and pitch accuracy and matches tagging, so the two tokens are complementary rather than redundant summaries.
  • The CPSD loss induces circle-of-fifths structure in a transformer that is not architecturally transposition-equivariant, meaning equivariance can be acquired through self-supervision rather than built into the architecture.
  • Sequence tokens from the shared backbone also improve over both monotask baselines on beat and chord, and over MERT on chord, indicating multitask benefits extend beyond the class tokens.

Reading between the lines

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

  • Beyond the paper, the same recipe should support adding extra class tokens for other pretexts (tempo-shift equivariance, masked prediction), since the two tested losses already share one backbone without interference; each new token would add a representation axis at no parameter cost.
  • A testable extension is to replace the plain average of the two tokens with a learned fusion (weighted sum or cross-attention) on the two tasks where averaging was not best, instrument recognition and beat tracking.
  • Because the CPSD guarantee formally assumes a one-hot pitch-class output, the strongest signal of this paper is on tonal music; an out-of-distribution check on atonal or percussive audio would tell how far the equivariant token generalizes.
  • The multi-token trick is not music-specific: in speech or bioacoustics, one token could be trained with a contrastive speaker objective and another with an equivariant shift objective, giving a compact multitask representation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes MT2, a ViT-1D architecture with two class tokens trained jointly on two self-supervised pretext tasks: a contrastive NT-Xent loss on one token and a CPSD equivariant loss on the other. The authors evaluate the frozen representations with a single linear probe on four global MIR tasks (tagging, instrument recognition, key estimation, pitch estimation) and two local tasks (beat tracking, chord estimation), comparing against single-token monotask ViT models and the 95M-parameter MERT model. They report that MT2 outperforms the monotask baselines and MERT's last-layer representation on most tasks, despite having 18x fewer parameters, and present a PCA analysis showing specialization of the two tokens to pitch-related and timbre-related structure, respectively.

Significance. If the central claim holds, the paper offers a parameter-efficient self-supervised recipe that combines contrastive and equivariant objectives in one small transformer, a practically useful result for MIR. The paper is commendable for using standard public downstream benchmarks, for releasing code and model weights, and for probing both models under the same protocol rather than relying on task-specific tuning. The qualitative PCA ring structure is a nice interpretability result that connects the equivariant token to the circle of fifths. However, the main attribution of the performance gains to the multi-class-token multitask mechanism is currently confounded with an input-representation difference, and several empirical results lack error bars or adequate test-set size; these issues need to be resolved before the central claim is fully supported.

major comments (3)
  1. [§2.5, Table 1] The central comparison is confounded by the input representation. MT2 defines its input patch sequence as xp = xcqt + xmel (§2.5), whereas ViT-Mel sees only xmel (§2.1) and ViT-CQT sees only xcqt (§2.3). Therefore, every MT2-versus-baseline difference in Table 1 can be attributed, in whole or in part, to the strictly richer input stream rather than to the multi-class-token multitask mechanism. The discussion in §4.1 interprets the gains as 'beneficial inductive sharing' without isolating this factor. I request an ablation: for example, a single-class-token monotask model trained on xcqt + xmel with only the contrastive loss or only the equivariant loss, or an MT2 variant in which both class tokens receive only one of the two representations. Without this experiment, the paper's headline claim that multitask multi-token learning is responsible for the gains is not established.
  2. [§2.4, Fig. 2] The theoretical guarantee for the equivariant loss is stated in §2.4 as: Dk is zero 'if and only if zequiv A and zequiv B have a single nonzero entry and differ by a circular shift of k semitones.' In practice, zequiv is a softmax output over learned chroma-like representations and will rarely be one-hot, especially for non-tonal audio. The circle-of-fifths structure is therefore only approximately enforced, and the PCA ring in Fig. 2 is qualitative evidence rather than a quantitative check. The paper should report the sparsity or entropy of zequiv (e.g., the mean effective number of active chroma classes on a validation set) or provide an ablation that sharpens or thresholds zequiv, so that the claimed key and pitch gains can be tied to the stated mechanism.
  3. [Tables 1 and 2] All reported numbers are single point estimates with no error bars or multiple seeds, and the chord experiment uses only 124 songs in total with an 8:1:1 split, so the chord-test comparison rests on roughly a dozen test songs. For a multi-dataset claim as strong as 'outperforms MERT on all tasks except beat tracking,' the main MT2-versus-baseline and MT2-versus-MERT comparisons should be repeated across several seeds (or bootstrap resamples of the test sets) and reported with mean plus/minus standard deviation. Without this, fine-grained differences such as MT2 Cont. 0.390 vs. MT2 Avg. 0.388 mAP on tagging, or MT2 Seq. 0.447 vs. ViT-CQT Seq. 0.542 on chord estimation, cannot be interpreted reliably.
minor comments (4)
  1. [Figure 1 caption] The caption contains a typo: 'mult-class-token' should be 'multi-class-token'.
  2. [§5] In the conclusion, 'this may due to the sequence tokens' should read 'this may be due to the sequence tokens.'
  3. [Abstract and §4.2] The claim that MT2 'outperforms MERT' is specifically about probing the last layer with a single linear layer, no dropout, and no hidden dimensions, whereas the MERT paper uses task-specific layer weighting and a larger probe. The paper discloses this in §4.2, but the abstract and conclusion should carry the qualifier 'under a same-protocol final-layer linear probe' so that readers do not overgeneralize the comparison.
  4. [§3.3] For the chord estimation experiment, please report the number of test frames and the per-class distribution, since a small, imbalanced test set of 25 classes can make a few frames dominate the accuracy difference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MT2's gains are measured on external benchmarks and the multi-token multitask claim is an empirical finding, not an identity.

full rationale

The central claim is that a single ViT-1D with two class tokens trained simultaneously on NT-Xent and CPSD losses outperforms single-token monotask models and the MERT baseline. This claim is tested on independent public datasets (MagnaTagATune, TinySOL, FMAKv2/GiantSteps, Ballroom/GTZAN, RWC/SWD) with frozen backbones and a single linear probe, so it is not a fitted parameter renamed as a prediction. Pretext hyperparameters (tau=0.1, omega=7, k in [-5,6]) come from prior work or are hand-set, and downstream results are not used to select them. The equivariant loss derives from the authors' prior STONE and S-key work, but those are prior, externally validated methods rather than a self-referential uniqueness argument, and the paper's one-hot condition in Section 2.4 is a stated mathematical property, not a suppressed premise. The skeptical point that MT2 consumes xp = xcqt + xmel while each baseline sees only one representation is a genuine experimental confound, but it is an internal-validity limitation, not circular reasoning: the comparison is not equivalent to its inputs by construction. The one-hot caveat likewise weakens the theoretical guarantee without making the derivation self-feeding. Because no load-bearing step reduces to its own output, the circularity score is 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claims rest on standard SSL assumptions (positive and negative pairing, transposition equivariance, near-one-hot chroma) plus a set of hand-set hyperparameters. No new physical or mathematical entity is postulated; the two class tokens are architectural components.

free parameters (8)
  • NT-Xent temperature tau = 0.1
    Hand-set temperature in Equation (2); scales all contrastive logits and affects hard-negative behavior.
  • CPSD circular frequency omega = 7
    Equation (4); chosen as coprime to 12 to align with the circle of fifths; basis of the equivariant objective.
  • Pitch-shift interval k sampling = uniform in [-5, +6]
    Range over which segment C is transposed; defines the equivariant pretext distribution.
  • CQT resolution = Q=12 bins/octave, J=8 octaves
    Input representation for the equivariant token; harmonic granularity and range are hand-set.
  • Mel spectrogram parameters = 128 bins, 31.5 Hz frame rate
    Input representation for the contrastive token; standard choice, not fitted to downstream tasks.
  • Transformer architecture = de=192, depth=12, heads=3
    Smallest ViT variant; capacity choice likely affects multitask trade-offs.
  • Training schedule = batch 128, LR 1e-4 to 5e-7 cosine, 600 epochs, 512 steps/epoch
    Chosen schedule; no ablation showing sensitivity.
  • Projection dimension for contrastive head = 512
    Head output dimension in Section 2.1; hand-set.
assumptions (6)
  • domain assumption A musical piece does not modulate between the two disjoint segments A and B (Section 2.4).
    The D0 CPSD term is minimized under the premise that tonality is constant within each pair of segments; if modulation occurs, the pretext target is ill-posed.
  • domain assumption Artificially transposing segment A by k semitones yields segment C whose key is exactly k semitones apart (Section 2.4).
    Equivariant targets Dk rely on pitch shifts preserving all other musical content; this is approximate for audio with vocals or non-pitched instruments.
  • standard math The softmax output zequiv is close to one-hot, so Dk zero implies a unique pitch class and circular shift (Section 2.4).
    The proof of Dk's zero condition uses the property of softmax and coprimality; real chroma vectors are not one-hot, so this is an idealization.
  • domain assumption Two segments from the same song form a positive pair and segments from other songs are valid negatives (Section 2.2).
    Contrastive learning's semantics depend on this assumption; false positives, such as the same song but very different sections, could hurt the contrastive token.
  • domain assumption Averaging the two class tokens is a useful fusion for downstream probes (Section 2.6).
    The paper hypothesizes complementarity and tests it empirically; there is no theoretical guarantee that averaging is optimal.
  • domain assumption Backpropagating both losses through shared attention and MLP weights does not destructively interfere (Section 2.5).
    The multitask architecture assumes gradient compatibility; the paper provides empirical evidence but no analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Class-Token Transformer for Multitask Self-supervised Music Information Retrieval." pith.science (2026). https://pith.science/paper/IYAX33LU

@misc{pith2026250712996,
  author       = {Pith},
  title        = {Pith review of: Multi-Class-Token Transformer for Multitask Self-supervised Music Information Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IYAX33LU}},
  note         = {Machine review of arXiv:2507.12996}
}
read the original abstract

Contrastive learning and equivariant learning are effective methods for self-supervised learning (SSL) for audio content analysis. Yet, their application to music information retrieval (MIR) faces a dilemma: the former is more effective on tagging (e.g., instrument recognition) but less effective on structured prediction (e.g., tonality estimation); The latter can match supervised methods on the specific task it is designed for, but it does not generalize well to other tasks. In this article, we adopt a best-of-both-worlds approach by training a deep neural network on both kinds of pretext tasks at once. The proposed new architecture is a Vision Transformer with 1-D spectrogram patches (ViT-1D), equipped with two class tokens, which are specialized to different self-supervised pretext tasks but optimized through the same model: hence the qualification of self-supervised multi-class-token multitask (MT2). The former class token optimizes cross-power spectral density (CPSD) for equivariant learning over the circle of fifths, while the latter optimizes normalized temperature-scaled cross-entropy (NT-Xent) for contrastive learning. MT2 combines the strengths of both pretext tasks and outperforms consistently both single-class-token ViT-1D models trained with either contrastive or equivariant learning. Averaging the two class tokens further improves performance on several tasks, highlighting the complementary nature of the representations learned by each class token. Furthermore, using the same single-linear-layer probing method on the features of last layer, MT2 outperforms MERT on all tasks except for beat tracking; achieving this with 18x fewer parameters thanks to its multitasking capabilities. Our SSL benchmark demonstrates the versatility of our multi-class-token multitask learning approach for MIR applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 39 canonical work pages

  1. [1]

    A cookbook of self- supervised learning,

    R. Balestriero, M. Ibrahim, V . Sobal, A. Morcos, S. Shekhar, T. Goldstein, F. Bordes, A. Bardes, G. Mialon, Y . Tian et al. , “A cookbook of self- supervised learning,” arXiv preprint arXiv:2304.12210 , 2023

    arXiv 2023

  2. [2]

    An experimental comparison of multi-view self-supervised methods for music tagging,

    G. Meseguer-Brocal, D. Desblancs, and R. Hennequin, “An experimental comparison of multi-view self-supervised methods for music tagging,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    work page 2024

  3. [3]

    Contrastive learning of musical representations,

    J. Spijkervet and J. A. Burgoyne, “Contrastive learning of musical representations,” in Proc. International Society for Music Information Retrieval (ISMIR) Conference , 2021

    work page 2021

  4. [4]

    Supervised and unsupervised learning of audio representations for music understanding,

    M. C. McCallum, F. Korzeniowski, S. Oramas, F. Gouyon, and A. F. Ehmann, “Supervised and unsupervised learning of audio representations for music understanding,” in Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2022

    work page 2022

  5. [5]

    Leave-one-equivariant: Allevi- ating invariance-related information loss in contrastive music represen- tations,

    J. Guinot, E. Quinton, and G. Fazekas, “Leave-one-equivariant: Allevi- ating invariance-related information loss in contrastive music represen- tations,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025, pp. 1–5

    work page 2025

  6. [6]

    Towards proper contrastive self- supervised learning strategies for music audio representation,

    J. Choi, S. Jang, H. Cho et al. , “Towards proper contrastive self- supervised learning strategies for music audio representation,” in Proc. IEEE International Conference on Multimedia and Expo (ICME) . IEEE, 2022, pp. 1–6

    work page 2022

  7. [7]

    J. S. Downie, A. F. Ehmann, M. Bay, and M. C. Jones, The music information retrieval evaluation exchange: Some observations and insights. Springer, 2010, pp. 93–115

    work page 2010

  8. [8]

    Zero-note samba: Self- supervised beat tracking,

    D. Desblancs, V . Lostanlen, and R. Hennequin, “Zero-note samba: Self- supervised beat tracking,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2023

    work page 2023

Show all 40 references
  1. [9]

    SPICE: Self-supervised pitch estimation,

    B. Gfeller, C. Frank, D. Roblek, M. Sharifi, M. Tagliasacchi, and M. Velimirovi´c, “SPICE: Self-supervised pitch estimation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 28, pp. 1118–1128, 2020

    2020

  2. [10]

    PESTO: Pitch estimation with self-supervised transposition-equivariant objective,

    A. Riou, S. Lattner, G. Hadjeres, and G. Peeters, “PESTO: Pitch estimation with self-supervised transposition-equivariant objective,” in Proc. International Society for Music Information Retrieval Conference (ISMIR), 2023

    2023

  3. [11]

    Equivariant self-supervision for musical tempo estimation,

    E. Quinton, “Equivariant self-supervision for musical tempo estimation,” in Proc. International Society for Music Information Retrieval Conference (ISMIR), 2022

    2022

  4. [12]

    Mert: Acoustic music understanding model with large-scale self-supervised training,

    Y . Li, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Lin, A. Ragni, E. Benetos, N. Gyenge, R. Dannenberg, R. Liu, W. Chen, G. Xia, Y . Shi, W. Huang, Y . Guo, and J. Fu, “Mert: Acoustic music understanding model with large-scale self-supervised training,” in Proc. Internatio...

    2023

  5. [13]

    A foundation model for music informatics,

    M. Won, Y .-N. Hung, and D. Le, “A foundation model for music informatics,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 1226–1230

    2024

  6. [14]

    Multi-tasking with joint semantic spaces for large-scale music annotation and retrieval,

    J. Weston, S. Bengio, and P. Hamel, “Multi-tasking with joint semantic spaces for large-scale music annotation and retrieval,” Journal of New Music Research, vol. 40, no. 4, pp. 337–348, 2011

    2011

  7. [15]

    Transfer learning in MIR: Sharing learned latent representations for music audio classification and similarity,

    P. Hamel, M. E. P. Davies, K. Yoshii, and M. Goto, “Transfer learning in MIR: Sharing learned latent representations for music audio classification and similarity,” in Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2013

    2013

  8. [16]

    The MIREX grand challenge: A framework of holistic user-experience evaluation in music information retrieval,

    X. Hu, J. H. Lee, D. Bainbridge, K. Choi, P. Organisciak, and J. S. Downie, “The MIREX grand challenge: A framework of holistic user-experience evaluation in music information retrieval,” in Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2014

    2014

  9. [17]

    Cross task study on MIREX recent results: An index for evolution measurement and some stagnation hypotheses,

    R. E. P. Scholz, G. L. Ramalho, and G. Cabral, “Cross task study on MIREX recent results: An index for evolution measurement and some stagnation hypotheses,” in Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2016

    2016

  10. [18]

    Multitask learning for frame- level instrument recognition,

    Y .-N. Hung, Y .-A. Chen, and Y .-H. Yang, “Multitask learning for frame- level instrument recognition,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2019, pp. 381–385

    2019

  11. [19]

    One deep music representation to rule them all? a comparative analysis of different representation learning strategies,

    J. Kim, J. Urbano, C. C. S. Liem, and A. Hanjalic, “One deep music representation to rule them all? a comparative analysis of different representation learning strategies,” Neural Computing and Applications , vol. 32, no. 4, pp. 1067–1093, 2020

    2020

  12. [20]

    Multitask learning based deep learning model for music artist and language recognition,

    Y . Singh and A. Biswas, “Multitask learning based deep learning model for music artist and language recognition,” in Proc. Workshop on Speech and Music Processing (SMP) , 2021

    2021

  13. [21]

    HEAR: Holistic evaluation of audio representations,

    J. Turian, J. Shier, H. R. Khan, B. Raj, B. W. Schuller, C. J. Steinmetz, C. Malloy, G. Tzanetakis, G. Velarde, K. McNally et al., “HEAR: Holistic evaluation of audio representations,” in Proceeedings of the Neural Information Processing Systems Conference (NeurIPS), Competiti...

    2022

  14. [22]

    Emergent musical properties of a transformer under contrastive self-supervised learning,

    Y . Kong, G. Meseguer-Brocal, V . Lostanlen, M. Lagrange, and R. Hen- nequin, “Emergent musical properties of a transformer under contrastive self-supervised learning,” International Society for Music Information Retrieval Conference (ISMIR 2025) , 2025

    2025

  15. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” in Proc. International Conference on Le...

    2021

  16. [24]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proc. IEEE/CVF International Conference on Computer Vision (ICCV) , 2021, pp. 9650–9660

    2021

  17. [25]

    STONE: Self-supervised tonality estimator,

    Y . Kong, V . Lostanlen, G. Meseguer-Brocal, S. Wong, M. Lagrange, and R. Hennequin, “STONE: Self-supervised tonality estimator,” Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2024

    2024

  18. [26]

    S-key: Self-supervised learning of major and minor keys from audio,

    Y . Kong, G. Meseguer-Brocal, V . Lostanlen, M. Lagrange, and R. Hen- nequin, “S-key: Self-supervised learning of major and minor keys from audio,” in ICASSP 2025 - IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025, pp. 1–5

    2025

  19. [27]

    Deep salience representations for f0 estimation in polyphonic music,

    R. Bittner, B. McFee, J. Salamon, P. Li, and J. Bello, “Deep salience representations for f0 estimation in polyphonic music,” in Proc. 18th International Society for Music Information Retrieval Conference (ISMIR) , Suzhou, China, Oct. 2017

    2017

  20. [28]

    Vision transformers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski, “Vision transformers need registers,” in Proc. International Conference on Learning Represen- tations (ICLR) , 2024

    2024

  21. [29]

    Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation,

    L. Xu, M. Bennamoun, F. Boussaid, H. Laga, W. Ouyang, and D. Xu, “Mctformer+: Multi-class token transformer for weakly supervised semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 8380–8395, 2024

    2024

  22. [30]

    Evaluation of algorithms using games: The case of music tagging,

    E. Law, K. West, M. I. Mandel, M. Bay, and J. S. Downie, “Evaluation of algorithms using games: The case of music tagging,” in Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2009, pp. 213–218

    2009

  23. [31]

    Sample-level deep convolutional neural networks for music auto-tagging using raw waveforms,

    J. Lee, J. Park, K. L. Kim, and J. Nam, “Sample-level deep convolutional neural networks for music auto-tagging using raw waveforms,” in Proc. International Conference on Sound and Music Computing (SMC) , 2017

    2017

  24. [32]

    Orchideasol: A dataset of extended instrumental techniques for computer- aided orchestration,

    C. E. Cella, D. Ghisi, V . Lostanlen, F. L ´evy, J. Fineberg, and Y . Maresz, “Orchideasol: A dataset of extended instrumental techniques for computer- aided orchestration,” in Proc. International Computer Music Conference (ICMC), 2020

    2020

  25. [33]

    Fmak: A dataset of key and mode annotations for the free music archive — extended abstract,

    S. Wong and G. Hernandez, “Fmak: A dataset of key and mode annotations for the free music archive — extended abstract,” in Proc. 24th International Society for Music Information Retrieval (ISMIR) Late- Breaking/Demo Papers, Milan, Italy, 2023

    2023

  26. [34]

    Fma: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “Fma: A dataset for music analysis,” Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2017

    2017

  27. [35]

    Two datasets for tempo estimation and key detection in electronic dance music annotated from user corrections,

    P. Knees, ´Angel Faraldo, P. Herrera, R. V ogl, S. B ¨ock, F. H ¨orschl¨ager, and M. L. Goff, “Two datasets for tempo estimation and key detection in electronic dance music annotated from user corrections,” in Proc. International Society for Music Information Retrieval Confere...

    2015

  28. [36]

    mir eval: A transparent implementation of common mir metrics

    C. Raffel, B. McFee, E. J. Humphrey, J. Salamon, O. Nieto, D. Liang, D. P. Ellis, and C. C. Raffel, “mir eval: A transparent implementation of common mir metrics.” in Proc. International Society for Music Information Retrieval Conference (ISMIR) , vol. 10, 2014, p. 2014

    2014

  29. [37]

    A review of rhythm description systems,

    F. Gouyon and S. Dixon, “A review of rhythm description systems,” in Proc. International Society for Music Information Retrieval Conference (ISMIR), 2004

    2004

  30. [38]

    Gtzan-rhythm: Extending the gtzan test-set with beat, downbeat and swing annotations,

    U. Marchand, Q. Fresnel, and G. Peeters, “Gtzan-rhythm: Extending the gtzan test-set with beat, downbeat and swing annotations,” Proc. Inter- national Society for Music Information Retrieval Late-breaking/Demo (ISMIR-LBD), 2015

    2015

  31. [39]

    Schubert winterreise dataset: A multimodal scenario for music analysis,

    C. Weiß, F. Zalkow, V . Arifi-M¨uller, M. M¨uller, H. V . Koops, A. V olk, and H. G. Grohganz, “Schubert winterreise dataset: A multimodal scenario for music analysis,” Journal on Computing and Cultural Heritage (JOCCH) , vol. 14, no. 2, pp. 1–18, 2021

    2021

  32. [40]

    RWC music database: Popular, classical, and jazz music databases,

    M. Goto and H. Hashiguchi, “RWC music database: Popular, classical, and jazz music databases,” Proc. International Society for Music Information Retrieval Conference (ISMIR) , 2002

    2002

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.