Pith. sign in

REVIEW 3 major objections 4 minor 64 references

Universal Music Representations? Evaluating Foundation Models on World Music Corpora

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

Pith's one-line read Audio foundation models are measurably Western-centric: their music-tagging accuracy declines on Greek, Turkish, and Indian classical corpora, and their advantage over a small baseline vanishes in few-shot settings.

desk verdict Worth a serious look if you do MIR fairness, but the cross-cultural gradient claim is more fragile than the paper lets on. read the letter →

arxiv 2506.17055 v1 pith:FCDTKKCH submitted 2025-06-20 cs.SD cs.IRcs.LGeess.AS

classification cs.SDcs.IRcs.LGeess.AS
keywords audiofoundationmodelsworldmusiccross-culturalevaluationtaggingfew-shotlearningLC-ProtonetsWestern-centricbiasrepresentation
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

This paper asks whether audio foundation models truly produce universal music representations. It probes five models on six corpora spanning Western, Greek, Turkish, and Indian classical traditions, using three setups: frozen-feature probing, targeted fine-tuning, and multi-label few-shot learning. The central finding is a consistent cultural gradient: tagging accuracy is highest on Western music and declines with cultural distance, with Indian classical corpora the hardest. Larger models generally do better on non-Western music, and the paper's setups reach state-of-the-art results on five of six datasets. Yet in low-resource few-shot conditions, foundation models perform only on par with a much smaller supervised baseline, which the authors read as evidence of implicit Western bias.

What carries the argument

The evaluation framework is the machinery: three complementary reading-outs of the same models. Probing trains a single-hidden-layer MLP on frozen embeddings; supervised fine-tuning (SFT) unfreezes only the last 1-2 layers; multi-label few-shot learning (ML-FSL) uses Label-Combination Prototypical Networks (LC-Protonets), which build a prototype per label combination and classify a query by nearest prototype. The cross-cultural lens comes from aligning six corpora to comparable durations and tag sets, so score gaps by tradition are interpreted as cultural distance. The LC-Protonets optimization collapses identical prototypes into a dictionary, preserving exact outputs while cutting inference from roughly two seconds to under 20 ms at 60 labels.

What would settle it

Run the same tagging setup on a matched Western "hard" dataset, such as dense solo instruments or non-tonal contemporary Western music, and compare the gap; if the cultural gradient persists when label difficulty and recording quality are controlled, the Western-bias reading holds, and if it disappears, dataset difficulty rather than cultural distance explains the drop. More directly, fine-tune a model on Carnatic data alone and test on Western corpora: a symmetric cultural penalty would be expected if pre-training data, not inherent musical distance, drives the asymmetry.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that current audio foundation models encode a measurable Western-centric bias. Across all models, tagging accuracy falls consistently by cultural distance, from roughly 90-97% ROC-AUC on Western corpora to 61-83% on Carnatic and Hindustani, with the music-only CLAP model showing the largest drop. Fine-tuning the last layers recovers 1-2 AUC points on average and helps the weakest model most, but does not consistently beat simple probing. In few-shot learning on extended tag sets, the best foundation model matches the much smaller VGG-ish feature extractor rather than surpassing it, and only on Western datasets does the foundation model clearly win. The paper also reports an optimized implementation of LC-Protonets that computes distances only to unique label-combination prototypes, giving 10-100x inference speedups with identical classification results.

Load-bearing premise

The cultural-distance reading assumes the six corpora and their tag sets are comparable proxies for musical traditions, so that a score drop on Indian classical music measures Western bias rather than dataset difficulty, tag quality, or recording conditions; the paper does not independently validate this comparability.

Editorial extensions

If this is right

  • Foundation model representations are not yet universal: their music-tagging accuracy degrades steadily with cultural distance from pre-training data.
  • Fine-tuning only the last layers yields modest gains and does not consistently beat probing, suggesting most musical knowledge is already frozen in earlier layers.
  • In low-resource few-shot scenarios, foundation models lose their edge over small supervised baselines, so world-music collections cannot rely on generic pre-training alone.
  • The optimized LC-Protonets makes few-shot tagging practical for large label sets, a typical condition for world music datasets.
  • State-of-the-art results on five of six datasets show that current models are strong on Western music and moderately transferable to related traditions, establishing a baseline benchmark for future progress.

Reading between the lines

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

  • The cultural gradient may partly reflect tag-set semantics: instrument and mood tags in the Indian classical corpora could be harder to infer from short clips, so a controlled experiment with matched tags would sharpen the bias claim.
  • The finding that SFT does not beat probing hints at a ceiling: because only the final layers are adapted, deeper layers retain Western priors, so fuller adaptation or low-rank adapters might close more of the gap than this constrained setup allows.
  • ML-FSL parity with VGG-ish suggests that few-shot evaluation is a more sensitive probe of representation universality than full-data probing and could become a standard fairness metric for audio foundation models.
  • Extending the framework to mode, makam, or raga recognition would test whether the bias is tag-specific or structural in the learned representations.
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 evaluates five audio foundation models (MERT-95M, MERT-330M, CLAP-Music, CLAP-Music&Speech, and Qwen2-Audio) plus a VGG-ish baseline on six music corpora spanning Western, Greek, Turkish, and Indian classical traditions, using three protocols: linear probing, targeted supervised fine-tuning, and multi-label few-shot learning. It reports declining tagging performance for culturally distant traditions, larger models generally performing better, state-of-the-art results on five of six datasets, and a computational optimization for LC-Protonets that reduces inference cost by up to 100x. The paper argues that these results demonstrate Western-centric bias in current audio foundation models.

Significance. If the central finding is established with appropriate controls, the paper would be a valuable benchmark for cross-cultural music tagging: it covers a diverse set of corpora, includes three complementary evaluation protocols, releases code, and proposes a practical few-shot optimization. The dataset and model coverage is broader than previous work, and the few-shot results offer a realistic low-resource scenario. However, the cross-cultural interpretation currently rests on an unvalidated comparability assumption, and the state-of-the-art claims need independent baselines before the contribution can be fully assessed.

major comments (3)
  1. [Section 3.2 and Table 2] The central cross-cultural bias conclusion is confounded by dataset difficulty. The VGG-ish baseline in Table 2 exhibits the same ROC-AUC gradient (91.23 on MagnaTagATune, 88.89 on FMA-medium, 80.97 on Lyra, 86.96 on Turkish-makam, 84.77 on Hindustani, 73.92 on Carnatic) even though VGG-ish is not a foundation model and does not share the claimed Western-centric pretraining bias. The paper does not report label statistics such as label density, number of positive labels per clip, or annotation quality, and the tag sets differ in size (50, 30, and 20 tags) and semantics across datasets. Therefore the aggregate cross-dataset comparisons in Section 5.1 cannot uniquely attribute the performance gradient to Western-centric pretraining. In fact, the relative advantage of the best foundation model over VGG-ish does not decline monotonically with cultural distance (e.g., Qwen2-Audio's ROC-AUC advantage over VGG-ish is -0.06 on MagnaTagATune, +5.47 on Lyra, -0.32 on Turkish-makam, and +8.30 on Carnatic), so a difficulty control is required before the main conclusion can be drawn.
  2. [Section 5.1 and Table 2] The claim of state-of-the-art performance on five of six datasets is self-referential: the previous SOTA values are taken from the authors' own prior work [28] for all datasets except MagnaTagATune, where an external system [52] is used. The contribution statement therefore reduces to 'we improve over our own earlier baselines' for most of the comparison. To substantiate the SOTA claim, the authors should compare against independent and current baselines on each dataset, or explicitly qualify the claim as an improvement over [28].
  3. [Section 4 and Tables 1-2] The supervised fine-tuning results are based on a single run with no error bars, which undermines the quantitative conclusions drawn from them. Section 5.1 states that SFT gives average gains of 1-2% in ROC-AUC, but several per-dataset SFT numbers are below the corresponding probing numbers (e.g., MERT-95M on Lyra: 84.89 vs. 85.61; CLAP-M&S on FMA-medium mAP: 57.11 vs. 59.13; MERT-330M on Carnatic: 77.05 vs. 77.73). Without repeated runs, these differences cannot be distinguished from noise. At minimum, the authors should report multiple SFT seeds for the key comparisons or temper the claims about SFT improvement and the SFT-versus-probing conclusions.
minor comments (4)
  1. [Section 3.3] The description of Qwen2-Audio feature extraction is ambiguous: 'last hidden state embeddings averaged across all layers of the whole model' should clarify whether the features come from the audio encoder or the full language model, and why a text prompt containing only the tags is used.
  2. [Section 4] The statement that representation extraction strategies and fine-tuned layer counts were 'optimized through preliminary experiments' should report the search space and validation results; otherwise cross-model comparisons may reflect tuning choices rather than intrinsic model quality.
  3. [Section 5.2] The comparison with the VGG-ish baseline should explicitly state that the VGG-ish numbers are taken from [29] and describe whether that baseline was trained on standard tag sets and then evaluated on extended tag sets; this is important for interpreting the few-shot comparisons.
  4. [Section 3.4 and Figure 3] The claim that the optimized LC-Protonets produces 'identical classification results' should be supported by an explicit equivalence statement (e.g., exact match of predictions) rather than only speed measurements, since the current wording implies a stronger guarantee than the experiments demonstrate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's evaluations are external benchmarks, and its self-citations are transparent baseline comparisons that do not force the central conclusions.

full rationale

The paper's central claims—that foundation models show varying cross-cultural generalization and that the authors' approaches reach state-of-the-art performance on five of six datasets—are empirical evaluations against fixed corpora and fixed tag sets, not quantities derived from fitted parameters or from the claims themselves. No equation in the paper defines a result in terms of the target conclusion; the LC-Protonets optimization in Section 3.4 is shown to be identical to the original method by a set-equality argument, and the reported speedups are measured. The main self-citations are the VGG-ish baseline numbers taken from [28,29] and the previous SOTA values from [28], but these are used transparently as external benchmarks for the same experimental setup, not as premises that entail the paper's conclusions. The Western-centric-bias interpretation is an inference from measured score gaps across datasets; it may be confounded by unmatched dataset difficulty, tag-set sizes, semantics, and label densities, but that is a validity or correctness concern, not circularity. Under the hard rules, no claim reduces to its inputs by construction, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on publicly available datasets and pretrained models, plus a handful of hand-chosen design decisions (layer counts, representation extraction, few-shot K). No new physical or mathematical entities are postulated. The main domain assumptions are that the corpora represent their cultures and that the VGG-ish literature baselines are comparable.

free parameters (5)
  • Fine-tuned layer counts per model = 1-2 layers (MERT-95M: 2, MERT-330M: 1, CLAP: 1 group, Qwen: 1)
    Chosen partly by RAM constraints; directly affects SFT and downstream ML-FSL scores.
  • Representation extraction strategy = Sum of last 4 MERT layers; CLAP audio projection; Qwen last hidden states
    Authors state these were optimized in preliminary experiments, so they are tuned choices rather than neutral defaults.
  • Few-shot support size K = 3 examples per label
    All ML-FSL experiments use K=3; performance is sensitive to this.
  • Probing MLP hidden size and learning rates = 512 units; lr=1e-3 (probing), lr=1e-4 (SFT)
    Standard hyperparameters but they were hand-chosen and affect all reported metrics.
  • Dataset duration balancing = Approx. 200 hours per corpus (Lyra 80h)
    Maximum audio durations were set to match dataset sizes, following [28]; this changes how much audio each model sees and could influence scores.
assumptions (5)
  • domain assumption The six corpora and their tag sets operationalize the musical traditions they are named after.
    Section 3.2 introduces Lyra as Greek folk, Turkish-makam, Hindustani and Carnatic as Indian classical; the cultural-distance interpretation of performance gaps depends on this.
  • domain assumption Pre-training data of the foundation models is Western-centric.
    Assumed from the literature, e.g., MERT-330M proprietary data 'with a strong Western bias' [6]; the paper does not measure the actual data distribution.
  • domain assumption Averaging ROC-AUC and mAP across datasets with different tag set sizes is a valid summary.
    Table 1 averages across datasets with 50, 30, or 20 tags, which weights large-tag sets more heavily in the aggregate.
  • domain assumption VGG-ish baseline numbers from [28,29] are reproducible and comparable to this setup.
    The paper uses literature results instead of re-running VGG-ish, relying on identical splits and preprocessing.
  • standard math Power-set prototype averaging is a valid similarity measure for multi-label few-shot classification.
    Eq. (1) follows the original LC-Protonets definition; no new mathematical foundation is introduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universal Music Representations? Evaluating Foundation Models on World Music Corpora." pith.science (2026). https://pith.science/paper/FCDTKKCH

@misc{pith2026250617055,
  author       = {Pith},
  title        = {Pith review of: Universal Music Representations? Evaluating Foundation Models on World Music Corpora},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCDTKKCH}},
  note         = {Machine review of arXiv:2506.17055}
}
read the original abstract

Foundation models have revolutionized music information retrieval, but questions remain about their ability to generalize across diverse musical traditions. This paper presents a comprehensive evaluation of five state-of-the-art audio foundation models across six musical corpora spanning Western popular, Greek, Turkish, and Indian classical traditions. We employ three complementary methodologies to investigate these models' cross-cultural capabilities: probing to assess inherent representations, targeted supervised fine-tuning of 1-2 layers, and multi-label few-shot learning for low-resource scenarios. Our analysis shows varying cross-cultural generalization, with larger models typically outperforming on non-Western music, though results decline for culturally distant traditions. Notably, our approaches achieve state-of-the-art performance on five out of six evaluated datasets, demonstrating the effectiveness of foundation models for world music understanding. We also find that our targeted fine-tuning approach does not consistently outperform probing across all settings, suggesting foundation models already encode substantial musical knowledge. Our evaluation framework and benchmarking results contribute to understanding how far current models are from achieving universal music representations while establishing metrics for future progress.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 52 canonical work pages

  1. [28]

    Sample-level CNN archi- tectures for music auto-tagging using raw waveforms,

    T. Kim, J. Lee, and J. Nam, “Sample-level CNN archi- tectures for music auto-tagging using raw waveforms,” in ICASSP. IEEE, 2018, pp. 366–370

  2. [52]

    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 ISMIR, 2009, pp. 387–392

  3. [1]

    universal language

    INTRODUCTION The notion of music as a “universal language” remains contested among scholars [1, 2]. While some musical elements transcend cultural boundaries, traditions have evolved with distinct characteristics and semantic content [3,4]. This tension between universality and cultural speci- ficity presents a complex challenge that modern artificial int...

  4. [2]

    RELATED WORK Foundation models. Foundation models for music have emerged by leveraging large-scale self-supervised or con- trastive learning on extensive audio datasets, enabling them to capture rich musical features applicable across di- verse tasks. Representative works include JukeMIR [11], which explored representations from the Jukebox genera- tive m...

  5. [3]

    As shown in Figure 1, we employ three complementary Figure 1

    METHODOLOGICAL FRAMEWORK Our methodological framework systematically evaluates whether foundation models can effectively represent mu- sical characteristics across diverse cultural traditions. As shown in Figure 1, we employ three complementary Figure 1 . Architectural overview of our evaluation framework showcasing three methodologies: (1) Probing (Prob....

  6. [4]

    We conducted 5 runs with different random seeds for both Probing and ML-FSL tasks, but a single run for SFT due to computational constraints

    EXPERIMENTAL SETUP Experiments and resources. We conducted 5 runs with different random seeds for both Probing and ML-FSL tasks, but a single run for SFT due to computational constraints. SFT trainable parameters varied: 14M for MERT-95M, 13M for MERT-330M, 25M for CLAP mod- els, and 56M for Qwen2-Audio. All experiments ran on an NVIDIA RTX A5000 GPU, and...

  7. [5]

    Overall, Qwen2-Audio achieves the highest performance with 88.59% ROC-AUC and 56.48% mAP in Probing, further improving to 89.37% ROC-AUC and 58.73% mAP after fine-tuning

    RESULTS 5.1 Probing and Supervised Fine-Tuning Table 1 presents the performance of the evaluated founda- tion models averaged across all datasets for both Probing and SFT tasks. Overall, Qwen2-Audio achieves the highest performance with 88.59% ROC-AUC and 56.48% mAP in Probing, further improving to 89.37% ROC-AUC and 58.73% mAP after fine-tuning. This is ...

  8. [6]

    Although these models achieved better performance than previous models for diverse music traditions, we found clear indi- cators of Western-centric bias

    CONCLUSIONS In this paper, we examined the universality of music repre- sentations in foundation models through a comprehensive methodological framework evaluating five state-of-the-art audio models across six world music corpora. Although these models achieved better performance than previous models for diverse music traditions, we found clear indi- cato...

Show all 64 references
  1. [7]

    We extract represen- tations from three different contexts: directly from the pre- trained model ( PT), from the hidden layer of the trained Figure 2

    that is detailed in subsection 3.4. We extract represen- tations from three different contexts: directly from the pre- trained model ( PT), from the hidden layer of the trained Figure 2 . Relationship between model size and per- formance, averaged over Probing and Supervised F...

  2. [8]

    Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,” CoRR, vol. abs/2311.07919, 2023

  3. [9]

    A foundation model for music informatics,

    M. Won, Y . Hung, and D. Le, “A foundation model for music informatics,” in ICASSP. IEEE, 2024, pp. 1226–1230

  4. [10]

    Foundation models for music: A survey,

    Y . Ma, A. Øland, A. Ragni, B. M. Del Sette, C. Saitis, C. Donahue, C. Lin, C. Plachouras, E. Benetos, E. Quinton et al. , “Foundation models for music: A survey,”CoRR, vol. abs/2408.14340, 2024

  5. [11]

    Codified au- dio language modeling learns useful representations for music information retrieval,

    R. Castellon, C. Donahue, and P. Liang, “Codified au- dio language modeling learns useful representations for music information retrieval,” in ISMIR, 2021, pp. 88–96

  6. [12]

    ACKNOWLEDGMENTS We would like to thank the reviewers for their valuable and constructive feedback, which helped us improve our study. This work has been partially supported by project MIS 5154714 of the National Recovery and Resilience Plan Greece 2.0 funded by the European Un...

  7. [13]

    Uni- versality and diversity in human song,

    S. A. Mehr, M. Singh, D. Knox, D. M. Ketter, D. Pickens-Jones, S. Atwood, C. Lucas, N. Jacoby, A. A. Egner, E. J. Hopkins, R. M. Howardet al., “Uni- versality and diversity in human song,” Science, vol. 366, 2019

  8. [14]

    Sta- tistical universals reveal the structures and functions of human music,

    P. E. Savage, S. Brown, E. Sakai, and T. E. Currie, “Sta- tistical universals reveal the structures and functions of human music,” Proceedings of the National Academy of Sciences, vol. 112, pp. 8987 – 8992, 2015

  9. [15]

    Cross-cultural perspectives on music and musicality,

    S. E. Trehub, J. Becker, and I. Morley, “Cross-cultural perspectives on music and musicality,” Philosophical Transactions of the Royal Society B: Biological Sci- ences, vol. 370, 2015

  10. [16]

    Narratives imagined in response to instrumental music reveal culture- bounded intersubjectivity,

    E. H. Margulis, P. C. M. Wong, C. Turnbull, B. M. Kubit, and J. D. McAuley, “Narratives imagined in response to instrumental music reveal culture- bounded intersubjectivity,” Proceedings of the Na- tional Academy of Sciences of the United States of America, vol. 119, 2022

  11. [17]

    On the opportunities and risks of foun- dation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. B. Alt- man, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, E. Brynjolfsson, S. Buch, D. Card, R. Castellon, N. S. Chatterji, A. S. Chen, K. Creel et al., “On the opportunities and risks of foun- dation model...

  12. [18]

    MERT: acoustic music understanding model with large-scale self-supervised training,

    Y . Li, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetos, N. Gyenge et al. , “MERT: acoustic music understanding model with large-scale self-supervised training,” in ICLR. OpenReview.net, 2024

  13. [19]

    Large-scale contrastive language- audio pretraining with feature fusion and keyword-to- caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language- audio pretraining with feature fusion and keyword-to- caption augmentation,” in ICASSP. IEEE, 2023, pp. 1–5

  14. [20]

    musicnn: Pre-trained convolu- tional neural networks for music audio tagging,

    J. Pons and X. Serra, “musicnn: Pre-trained convolu- tional neural networks for music audio tagging,”CoRR, vol. abs/1909.06654, 2019

  15. [21]

    AST: audio spec- trogram transformer,

    Y . Gong, Y . Chung, and J. R. Glass, “AST: audio spec- trogram transformer,” inInterspeech. ISCA, 2021, pp. 571–575

  16. [22]

    Computational analysis of world music corpora,

    M. Panteli, “Computational analysis of world music corpora,” Ph.D. dissertation, Queen Mary University of London, UK, 2018

  17. [23]

    Automatic makam recognition using chroma features,

    E. Demirel, B. Bozkurt, and X. Serra, “Automatic makam recognition using chroma features,” in 8th In- ternational Workshop on F olk Music Analysis , 2018, pp. 19–24

  18. [24]

    Jukebox: A generative model for music,

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

  19. [25]

    Supervised and un- supervised learning of audio representations for music understanding,

    M. C. McCallum, F. Korzeniowski, S. Oramas, F. Gouyon, and A. F. Ehmann, “Supervised and un- supervised learning of audio representations for music understanding,” in ISMIR, 2022, pp. 256–263

  20. [26]

    Map-music2vec: A simple and ef- fective baseline for self-supervised music audio repre- sentation learning,

    Y . Li, R. Yuan, G. Zhang, Y . Ma, C. Lin, X. Chen, A. Ragni, H. Yin, Z. Hu, H. He, E. Benetos, N. Gyenge, R. Liu, and J. Fu, “Map-music2vec: A simple and ef- fective baseline for self-supervised music audio repre- sentation learning,” CoRR, vol. abs/2212.02508, 2022

  21. [27]

    Deep neural networks for music tagging,

    K. Choi, “Deep neural networks for music tagging,” Ph.D. dissertation, Queen Mary University of London, UK, 2018

  22. [29]

    This approach extends Proto- typical Networks [30] by creating prototypes for each la- bel combination, rather than generating one prototype per label

    for few-shot learning. This approach extends Proto- typical Networks [30] by creating prototypes for each la- bel combination, rather than generating one prototype per label. While established benchmarks for evaluating repre- sentations on downstream tasks typically employ pro...

  23. [30]

    Eval- uation of cnn-based automatic music tagging models,

    M. Won, A. Ferraro, D. Bogdanov, and X. Serra, “Eval- uation of cnn-based automatic music tagging models,” CoRR, vol. abs/2006.00751, 2020

  24. [31]

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

    J. Lee, J. Park, K. L. Kim, and J. Nam, “Sample- level deep convolutional neural networks for mu- sic auto-tagging using raw waveforms,” CoRR, vol. abs/1703.01789, 2017

  25. [32]

    CNN architectures for large-scale audio classification,

    S. Hershey, S. Chaudhuri, D. P. W. Ellis, J. F. Gem- meke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seybold et al., “CNN architectures for large-scale audio classification,” inICASSP. IEEE, 2017, pp. 131–135

  26. [33]

    MARBLE: music audio representation benchmark for universal evalua- tion,

    R. Yuan, Y . Ma, Y . Li, G. Zhang, X. Chen, H. Yin, L. Zhuo, Y . Liu, J. Huang, Z. Tian, B. Deng, N. Wang, C. Lin, E. Benetos, A. Ragni et al., “MARBLE: music audio representation benchmark for universal evalua- tion,” in NeurIPS, 2023

  27. [34]

    HTS-AT: A hierarchical token- semantic audio transformer for sound classification and detection,

    K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dubnov, “HTS-AT: A hierarchical token- semantic audio transformer for sound classification and detection,” in ICASSP. IEEE, 2022, pp. 646–650

  28. [35]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in ICCV. IEEE, 2021, pp. 9992–10 002

  29. [36]

    Cri- tiquing task- versus goal-oriented approaches: A case for makam recognition,

    K. K. Ganguli, S. Sentürk, and C. Guedes, “Cri- tiquing task- versus goal-oriented approaches: A case for makam recognition,” in ISMIR, 2022, pp. 369–376

  30. [37]

    Classification of indian clas- sical music with time-series matching deep learning approach,

    A. K. Sharma, G. Aggarwal, S. Bhardwaj, P. Chakrabarti, T. Chakrabarti, J. H. Abawajy, S. Bhattacharyya et al., “Classification of indian clas- sical music with time-series matching deep learning approach,” IEEE Access, vol. 9, pp. 102 041–102 052, 2021

  31. [38]

    KDC: an open corpus for computational research of dastg ¯ahi music,

    B. Nikzat and R. C. Repetto, “KDC: an open corpus for computational research of dastg ¯ahi music,” in ISMIR, 2022, pp. 321–328

  32. [39]

    For world music traditions, we incorporate the Lyra dataset

    (25,863 clips) and FMA-medium [40] (25,000 tracks). For world music traditions, we incorporate the Lyra dataset

  33. [40]

    Finding tori: Self-supervised learning for analyzing korean folk song,

    D. Han, R. C. Repetto, and D. Jeong, “Finding tori: Self-supervised learning for analyzing korean folk song,” in ISMIR, 2023, pp. 440–447

  34. [41]

    Following [28], we set maximum audio durations to achieve similar sizes between datasets and prepare their metadata for the auto-tagging task

    with 1,570 recordings of Greek folk music, and three collections from the CompMusic project [42]: the Turkish- makam corpus [43, 44] (5,297 recordings) as well as Hin- dustani [45] (1,204 recordings) and Carnatic [45] (2,612 recordings) of Indian classical music. Following [28...

  35. [42]

    From west to east: Who can understand the music of the others better?

    C. Papaioannou, E. Benetos, and A. Potamianos, “From west to east: Who can understand the music of the others better?” in ISMIR, 2023, pp. 311–318

  36. [43]

    LC-Protonets: Multi-label few-shot learning for world music audio tagging,

    ——, “LC-Protonets: Multi-label few-shot learning for world music audio tagging,” IEEE Open Journal of Signal Processing, vol. 6, pp. 138–146, 2025

  37. [44]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. S. Zemel, “Prototypical networks for few-shot learning,” in NIPS, 2017, pp. 4077–4087

  38. [45]

    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. Ve- larde, K. McNally, M. Henry, N. Pinto, C. Noufi et al., “HEAR: holistic evaluation of audio representations,” in NeurIPS (Competition and Demos) , ser. Proceed- ings of M...

  39. [46]

    SUPERB: speech process- ing universal performance benchmark,

    S. Yang, P. Chi, Y . Chuang, C. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G. Lin, T. Huang, W. Tseng, K. Lee et al. , “SUPERB: speech process- ing universal performance benchmark,” inInterspeech. ISCA, 2021, pp. 1194–1198

  40. [47]

    Continual pre-training of large language models: How to (re)warm your model?

    K. Gupta, B. Thérien, A. Ibrahim, M. L. Richter, Q. Anthony, E. Belilovsky, I. Rish, and T. Lesort, “Continual pre-training of large language models: How to (re)warm your model?” CoRR, vol. abs/2308.04014, 2023

  41. [48]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Lin, C. Zhou, and J. Zhou, “Qwen2-audio technical report,” CoRR, vol. abs/2407.10759, 2024

  42. [49]

    We applied learning rate warmup and cosine scheduling for the first 5% of SFT epochs

    with identical β parameters but learning rate 10−4, model-specific batch sizes (to fit maximum available re- sources) with gradient accumulation to simulate batch size 16 across all setups, patience 5, and maximum 30 epochs. We applied learning rate warmup and cosine schedulin...

  43. [50]

    Attention is All you Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is All you Need,” inAdvances in Neural In- formation Processing Systems, 2017

  44. [51]

    Very deep convolu- tional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolu- tional networks for large-scale image recognition,” in ICLR, 2015

  45. [53]

    FMA: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “FMA: A dataset for music analysis,” in ISMIR, 2017, pp. 316–323

  46. [54]

    A dataset for greek traditional and folk music: Lyra,

    C. Papaioannou, I. Valiantzas, T. Giannakopoulos, M. A. Kaliakatsos-Papakostas, and A. Potamianos, “A dataset for greek traditional and folk music: Lyra,” in ISMIR, 2022, pp. 377–383

  47. [55]

    Creating research corpora for the compu- tational study of music: the case of the compmusic project,

    X. Serra, “Creating research corpora for the compu- tational study of music: the case of the compmusic project,” in Semantic Audio. Audio Engineering So- ciety, 2014

  48. [56]

    A corpus for computational research of turk- ish makam music,

    B. Uyar, H. S. Atli, S. Sentürk, B. Bozkurt, and X. Serra, “A corpus for computational research of turk- ish makam music,” inDLfM@JCDL. ACM, 2014, pp. 1–7

  49. [57]

    Computational analysis of audio record- ings and music scores for the description and discovery of ottoman-turkish makam music,

    S. Sentürk, “Computational analysis of audio record- ings and music scores for the description and discovery of ottoman-turkish makam music,” Ph.D. dissertation, Pompeu Fabra University, Spain, 2017

  50. [58]

    Corpora for music information research in indian art music,

    A. Srinivasamurthy, G. K. Koduri, S. Gulati, V . Ishwar, and X. Serra, “Corpora for music information research in indian art music,” in ICMC. Michigan Publishing, 2014

  51. [59]

    Overcom- ing catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. C. Rabinowitz, J. Ve- ness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell, “Overcom- ing catastrophic forgetting in neural networks,” CoRR, vol. abs/1612.00796, 2016

  52. [60]

    Adam: A method for stochas- tic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochas- tic optimization,” in ICLR, 2015

  53. [61]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” CoRR, vol. abs/1711.05101, 2019

  54. [62]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” inICLR. Open- Review.net, 2022

  55. [63]

    The relationship between precision-recall and roc curves,

    J. Davis and M. Goadrich, “The relationship between precision-recall and roc curves,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 233–240

  56. [64]

    Mulan: A joint embedding of music audio and natural language,

    Q. Huang, A. Jansen, J. Lee, R. Ganti, J. Y . Li, and D. P. W. Ellis, “Mulan: A joint embedding of music audio and natural language,” in ISMIR, 2022, pp. 559– 566

Pith tools

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