Pith. sign in

REVIEW 4 major objections 8 minor 1 cited by

Music for All: Representational Bias and Cross-Cultural Adaptability of Music Generation Models

T0 review · 4 major / 8 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Music-generation models are trained on a corpus that is about 94 percent Western, and adapter fine-tuning fixes only some under-represented genre/model pairs.

desk verdict The 5.7% headline is not supported by the paper's own appendix, but the adapter experiments and the honest reporting of their mixed results make this a worthwhile read and a worthy revision. read the letter →

arxiv 2502.07328 v3 pith:G6G27MJG submitted 2025-02-11 cs.SD cs.AIcs.CLcs.LGcs.MM

Atharva Mehta , Shivam Chauhan , Amirbek Djanibekov , Atharva Kulkarni , Gus Xia , Monojit Choudhury This is my paper · ORCID
classification cs.SDcs.AIcs.CLcs.LGcs.MM
keywords musicgenerationrepresentationalbiasnon-Westernparameter-efficientfine-tuningbottleneckadaptersHindustaniClassicalTurkishMakamcross-culturaltransfer
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 tries to establish two things: that existing music-generation datasets are overwhelmingly Western, and that lightweight adapter fine-tuning is only a partial remedy. The authors surveyed 152 dataset papers and report that about 94% of the roughly one million hours of music data come from Western traditions, with only 5.7% from South Asian, Middle Eastern, Oceanian, Central Asian, Latin American, and African music. They then added small bottleneck adapters to MusicGen and Mustango and fine-tuned on Hindustani Classical and Turkish Makam recordings. Human arena evaluations showed real but uneven gains: Mustango improved on Hindustani Classical and MusicGen on Turkish Makam, while other pairings, especially creative cross-genre prompts, regressed. If the finding holds, no single cheap fine-tuning recipe fixes cultural bias; collection of diverse data and base-model design are the real levers.

What carries the argument

The work is carried by two mechanisms. The first is the Bottleneck Residual Adapter, a small trainable block (for MusicGen, a linear compression, nonlinearity, and projection placed after the transformer decoder; for Mustango, convolution-based adapters inside the UNet after cross-attention blocks) that adds about 2 million parameters, or 0.1% of each model, while the base weights stay frozen. These adapters let the authors fine-tune on 23.24 hours of Hindustani Classical and 121.16 hours of Turkish Makam audio. The second is an arena-style human evaluation built on Bloom's taxonomy, which separates Recall (reproduce known combinations), Analysis (recombine known elements in novel ways), and Creativity (blend genres) and ranks models with ELO ratings per criterion.

What would settle it

Recount the paper's own Appendix Table 3 with East Asian hours (2,746.73) counted as non-Western: the non-Western share rises from 5.7% to about 32% of the 10,372.88 surveyed hours, which would overturn the abstract's census claim.

Watch

Extended reading notes

Core claim

The central claim is that text-to-music models inherit a heavily Western training corpus, and that parameter-efficient adaptation cannot be assumed to transfer across music cultures. On the data side, the paper reports that only 5.7% of the hours in the surveyed music-generation datasets come from non-Western regions, with Pop, Rock, and Western Classical dominating the genre distribution. On the adaptability side, adapter fine-tuning with less than 1% of trainable parameters is effective for specific pairs: Mustango's overall ELO rose by about 8% on Hindustani Classical and MusicGen's by about 4% on Turkish Makam, but the same recipe degrades other pairs, and in both genres the fine-tuned models scored worse than baselines on creativity queries. The authors read this as evidence that architecture, training data, and base-model design determine cross-cultural adaptability more than the fine-tuning method itself.

Load-bearing premise

The headline 5.7% figure assumes that the 152 dataset papers sampled from ten conferences stand in for all existing music-generation datasets and that East Asian recordings can be counted as part of the well-explored, effectively Western majority; the paper gives no explicit definition of 'Western.'

Editorial extensions

If this is right

  • If the surveyed datasets are representative, new text-to-music models trained on the same corpora will continue to generate non-Western prompts poorly, particularly for South Asian, Middle Eastern, African, and Latin American genres.
  • Adapter fine-tuning can raise human-judged quality for some under-represented genres without retraining the full model, using less than 1% of parameters and roughly ten to twelve hours of GPU training.
  • The same adapters can lower quality or creativity on other genre/model pairs, so parameter-efficient fine-tuning is not a reliable universal fix and should be evaluated per genre and per criterion.
  • Because fine-tuning tends to trade away cross-genre creativity, improving equality in music AI requires collecting diverse training data and designing base models for transfer rather than relying only on post-hoc adaptation.

Reading between the lines

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

  • If the 5.7% figure is sensitive to where East Asian music is placed, then a binary West/non-West taxonomy is too coarse; a regional or genre-family breakdown would give model builders a clearer target for data collection.
  • The paper's own ELO tables show fine-tuned models losing on creativity queries even when they win on recall and analysis, suggesting that adapter training may cause knowledge attrition; a testable extension is to mix frozen and adapter weights or add a small general-domain buffer.
  • The inverse results across genres suggest an interaction between adapter placement and a genre's melodic and rhythmic structure, which could be tested by applying the same two adapter recipes to Carnatic, Arab-Andalusian, and Beijing opera data already available in the same corpus.
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

4 major / 8 minor

Summary. The paper studies representational bias in text-to-music generation models. It reports a dataset audit of 152 dataset papers from 10 conferences, claiming that only 5.7% of the total hours of existing music datasets come from non-Western genres. It then performs adapter-based PEFT experiments on MusicGen and Mustango for Hindustani Classical and Turkish Makam music, evaluating the adapted models with objective metrics (FAD, FD, KLD, PSNR) and a Bloom's-Taxonomy-inspired arena-style human evaluation with ELO ratings. The qualitative conclusion is that PEFT adaptation is non-trivial: it helps some model/genre pairs and hurts others, so base-model design and data collection matter more than a universal fine-tuning fix.

Significance. If the quantitative audit were reliable, the headline 5.7% statistic would be an important evidence point for debates about cultural diversity in generative music systems. The PEFT experiments address a timely and practical question, and the release of code and model adapters is a useful contribution. The human evaluation framework — with inter-annotator agreement computed and reported — is a methodological strength, and the paper candidly states its limitations, including the possible unsuitability of the PANN-CNN14 backbone for non-Western genres. However, the central quantitative claim is currently not anchored to a consistent denominator and the definition of 'Western' is left implicit, so the audit's headline number is not yet a reliable scientific result. The qualitative finding that PEFT effectiveness varies across genres and models is better supported, but it does not by itself repair the audit's quantitative premise.

major comments (4)
  1. [§2.1.1, §2.2, Tables 3–4] The 5.7% headline is not derived from a consistent denominator. Appendix A Table 3 lists regional durations whose sum is 10,372.88 hours, while Table 4 — explicitly labeled '10^3 hours' — sums to about 1.07M hours, matching the '1 million+ hours' claim in §2.1.1. The 94%/5.7% split in §2.2 is computed from the Table 3 regional sums (Europe+America+East Asia ≈ 94.4%; the listed non-Western regions ≈ 5.6%), i.e., from a 10,372.88-hour base, not from the 1M+ hour corpus. If Table 3 entries are actually in thousands of hours, the regional total becomes ~10.4M hours, contradicting §2.1.1. The paper must reconcile the units and restate the headline percentage on a well-defined denominator.
  2. [§2.2, Appendix A Table 3] The category 'Western' is never defined, and East Asian data are silently absorbed into the 'Western' majority. Table 3 lists 2,746.73 hours of East Asian material under a separate row, yet §2.2 treats East Asia as part of the roughly 94% Western share. Since East Asian art music is not self-evidently Western, reclassifying these hours as non-Western changes the non-Western share from about 5.6% to about 32% of the tabulated corpus. The paper needs an explicit definition of the 'Western'/'non-Western' boundary and should report how the headline figure changes under plausible alternative classifications.
  3. [§4.1, Table 2] The claim that 'Both finetuned versions of MusicGen and Mustango demonstrate superior performance compared to their baseline counterparts across all evaluated metrics' is contradicted by Table 2. For Hindustani Classical, MGF's PSNR (16.18) is below MGB's (16.23), and MGF's FAD (40.04) is essentially unchanged from MGB's (40.05). For Turkish Makam, MTF's PSNR (16.17) is below MTB's (16.60). The objective metrics are single-point estimates with no confidence intervals or repeated-generation variability reported, so statements of 'significant improvement' in §4.1 are not supported as stated.
  4. [Limitations (after §5)] The limitations paragraph concedes that the backbone models used to compute FAD, FD, and KLD 'may not have been trained on various underrepresented genres, resulting in an erroneous portrayal of genres.' This limitation directly affects the interpretation of Table 2, and the paper should present the objective metrics as auxiliary evidence only, with the primary evidence for adaptation quality resting on the human ELO results. The qualitative conclusion that PEFT effectiveness varies across genres and models is supported by the human evaluation, but the objective-metric narrative in §4.1 needs to be weakened accordingly.
minor comments (8)
  1. [§4.2.3] The sentence 'MGB is better than MTF' appears to be a typo; Table 7 shows MTF (1577) above MGB (1525) for Hindustani Classical overall, and the surrounding text says MTF outperforms all models. This should read 'MGB is better than MGF' or 'MTF is better than MGB.'
  2. [Appendix A Table 3] The paper counts in Table 3 sum to 223, not the 152 dataset papers reported in §2.1.1; a footnote explaining that a paper can cover multiple regions would remove the apparent inconsistency.
  3. [§2.1.1] The sentence '7.9% of the datasets totaling 5,772 hours were excluded' is ambiguous: 5,772 hours is far less than 7.9% of 1M+ hours, so the percentage likely refers to the number of dataset papers, not the hours. Please rephrase to state both counts clearly.
  4. [Table 4] The note that 'Duration (Dur.) is represented as 10^3 hours' is easy to overlook and is inconsistent with Table 3, whose header reads 'Duration (hrs.)'. Use a uniform unit across all tables or explicitly state the multiplier in every table header.
  5. [Appendix G] The ELO computation uses a K-factor of 15, chosen because of the limited number of matches; this is a free parameter, and the paper should either report how the rating conclusions change for a range of K values or add confidence intervals around the ELO scores.
  6. [§3.1] The preprocessing section explains that 405 hours of Turkish Makam audio was collected via the Dunya API, but only 121.16 hours remained after preprocessing; the large reduction is not explained and should be.
  7. [Appendix I, §4.2.2] Appendix I lists four annotators (three from India and one from Uzbekistan), but §4.2.2 says only two annotators participated in phase 1; please clarify the total number of annotators and their roles in phases 1 and 2.
  8. [Throughout] There are several typos, e.g., 'Mustangto' in §4, 'underrerpresented' in §5, and 'ranking models on on their adherence' in the contributions list; these should be corrected in a final polish.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset audit is an independent measurement and the PEFT evaluation uses held-out test audio plus human comparisons; the only self-citation is non-load-bearing inspiration.

full rationale

The paper's two load-bearing claims are (1) that only 5.7% of music dataset hours come from non-Western genres, and (2) that PEFT-based adaptation has genre- and model-dependent effectiveness. Neither claim reduces to its own inputs by construction. The 5.7% figure comes from a newly performed manual audit of 152 dataset papers with stated inclusion and exclusion criteria; it is a tabulation of annotated hours, not a quantity defined in terms of the conclusion it supports. The PEFT experiments train adapters on one split of the genre datasets, freeze the base models, and evaluate on held-out test audio using both objective metrics (FAD, FD, KLD, PSNR) and human ELO comparisons; no fitted parameter is relabeled as a prediction, and no evaluation target is used as a training input. The only self-citation, Joshi et al. (2020) (which shares an author), is explicitly cited as methodological inspiration for conducting a survey, not as evidence for the 5.7% result or for the PEFT conclusions, so it is not load-bearing. Concerns raised about the audit—such as the lack of an explicit definition of 'Western', the apparent mismatch between the claimed '1 million+ hours' in Section 2.1.1 and the regional hours in Table 3, and the selective removal of some ELO comparisons—are validity or correctness risks rather than circularity, because they do not involve a derivation that is equivalent to its own input by definition. The paper is therefore self-contained with respect to circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper adds no new physical entities and fits no scientific constants. The ELO K factor is the main hand-set evaluation parameter. The four axioms above are the premises that the headline measurements and the human evaluation framework depend on.

free parameters (1)
  • ELO K-factor = 15
    Hand-set to 15 because the match count is small; a higher K would make each match move ratings more. It changes rating magnitudes but not the sign of preferences.
assumptions (4)
  • domain assumption The surveyed corpus of 152 dataset papers from 10 conferences is representative of 'existing music datasets.'
    The abstract generalizes to 'existing music datasets,' but Section 2.1 only searches 10 venues with keyword queries and excludes 7.9% of datasets for missing metadata.
  • ad hoc to paper East Asian music can be grouped with European and American music as part of the well-explored majority rather than the non-Western share.
    Section 2.2 treats European, East Asian, and American data together as the 'well explored' 84.5%, and the headline 5.7% excludes East Asian hours, without any stated definition of 'Western.'
  • domain assumption PANN-CNN14 audio features are adequate for computing FAD, FD, and KLD for Hindustani Classical and Turkish Makam music.
    Section 4.1 uses this backbone for all objective metrics; the Limitations section concedes the backbone may not have been trained on underrepresented genres, which could bias the metrics.
  • ad hoc to paper The Bloom's taxonomy levels Recall, Analysis, and Creativity are meaningful categories for evaluating text-to-music generation.
    Section 4.2.1 introduces the three levels as evaluation conditions without validation against music perception or music cognition research.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Music for All: Representational Bias and Cross-Cultural Adaptability of Music Generation Models." pith.science (2026). https://pith.science/paper/G6G27MJG

@misc{pith2026250207328,
  author       = {Pith},
  title        = {Pith review of: Music for All: Representational Bias and Cross-Cultural Adaptability of Music Generation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G6G27MJG}},
  note         = {Machine review of arXiv:2502.07328}
}
read the original abstract

The advent of Music-Language Models has greatly enhanced the automatic music generation capability of AI systems, but they are also limited in their coverage of the musical genres and cultures of the world. We present a study of the datasets and research papers for music generation and quantify the bias and under-representation of genres. We find that only 5.7% of the total hours of existing music datasets come from non-Western genres, which naturally leads to disparate performance of the models across genres. We then investigate the efficacy of Parameter-Efficient Fine-Tuning (PEFT) techniques in mitigating this bias. Our experiments with two popular models -- MusicGen and Mustango, for two underrepresented non-Western music traditions -- Hindustani Classical and Turkish Makam music, highlight the promises as well as the non-triviality of cross-genre adaptation of music through small datasets, implying the need for more equitable baseline music-language models that are designed for cross-cultural transfer learning.

Figures

Figures reproduced from arXiv: 2502.07328 by the authors.

Figure 1
Figure 1. The bottom left piechart shows the global distribution of genre. Each piechart in the map shows the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Mustango & MusicGen settings for low-resource fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Inter Annotator Agreement metrics for Hindustani Classical & Turkish Makam. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Scaled ELO ratings for each model in Hindustani Classical and Turkish Makam music. Categories [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Screenshots of Label Studio, annotation tool for Inter Annotator Agreement and ELO rating comparison [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Exploring Adapter Design Tradeoffs for Low Resource Music Generation

    cs.SD 2025-06 conditional novelty 5.0 of 10

    Adapter placement, architecture, and size strongly change generation quality and cost for MusicGen and Mustango on two non-Western genres, with late-layer, mid-sized (40M) adapters reported as the best tradeoff.

Reference graph

Works this paper leans on

47 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Andrea Agostinelli, Timo I Denk, Zal \'a n Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al. 2023. Musiclm: Generating music from text. arXiv preprint arXiv:2301.11325

    arXiv 2023

  4. [4]

    Armstrong

    P. Armstrong. 2010. Bloom’s taxonomy. https://cft.vanderbilt.edu/guides-sub-pages/blooms-taxonomy/. Retrieved October 12, 2024

    work page 2010

  5. [5]

    Ankur Bapna and Orhan Firat. 2019. https://doi.org/10.18653/v1/D19-1165 Simple, scalable adaptation for neural machine translation . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1538--1548, Hong Kong, China. Associat...

    doi:10.18653/v1/d19-1165 2019

  6. [6]

    Bender and Batya Friedman

    Emily M. Bender and Batya Friedman. 2018. https://doi.org/10.1162/tacl_a_00041 Data statements for natural language processing: Toward mitigating system bias and enabling better science . Transactions of the Association for Computational Linguistics, 6:587--604

    doi:10.1162/tacl_a_00041 2018

  7. [7]

    Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 610--623

    2021

  8. [8]

    Angelopoulos, Tianle Li, Dacheng Li, Banghua Zhu, Hao Zhang, Michael I

    Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios N. Angelopoulos, Tianle Li, Dacheng Li, Banghua Zhu, Hao Zhang, Michael I. Jordan, Joseph E. Gonzalez, and Ion Stoica. 2024. Chatbot arena: an open platform for evaluating llms by human preference. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org

    work page 2024

Show all 47 references
  1. [9]

    Cholewiak, Panos Ipeirotis, Victor Silva, and Arun Kannawadi

    Steven A. Cholewiak, Panos Ipeirotis, Victor Silva, and Arun Kannawadi. 2021. https://doi.org/10.5281/zenodo.5764801 Scholarly: Simple access to google scholar authors and citation using python

    2021 doi

  2. [10]

    Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinso...

    2024

  3. [11]

    Sheng-Kuan Chung. 2006. Digital storytelling in integrated arts education. The International Journal of Arts Education, 4(1):33--50

    2006

  4. [12]

    Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Defossez. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/94b472a1842cd7c56dcb125fb2765fbd-Paper-Conference.pdf Simple and controllable music generation . In A...

    2023

  5. [13]

    Kate Crawford. 2016. https://www.nytimes.com/2016/06/26/opinion/sunday/artificial-intelligences-white-guy-problem.html Artificial intelligence’s white guy problem . The New York Times. Accessed: 2024-10-16

    2016

  6. [14]

    Joe Davison. 2021. https://api.semanticscholar.org/CorpusID:235356070 Compacter: Efficient low-rank hypercomplex adapter layers . In Neural Information Processing Systems

    2021

  7. [15]

    Alexandre D \'e fossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. 2023. https://openreview.net/forum?id=ivCd8z8zR2 High fidelity neural audio compression . Transactions on Machine Learning Research. Featured Certification, Reproducibility Certification

    2023

  8. [16]

    Hao-Wen Dong, Wen-Yi Hsiao, Li-Chia Yang, and Yi-Hsuan Yang. 2018. Musegan: Multi-track sequential generative adversarial networks for symbolic music generation and accompaniment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32

    2018

  9. [17]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790--2799. PMLR

    2019

  10. [18]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 Lora: Low-rank adaptation of large language models . CoRR, abs/2106.09685

    2021 arXiv

  11. [19]

    Park, Tao Wang, Timo I

    Qingqing Huang, Daniel S. Park, Tao Wang, Timo I. Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jiahui Yu, Christian Havn Frank, Jesse Engel, Quoc V. Le, William Chan, and Wei Han. 2023. https://api.semanticscholar.org/CorpusID:256662408 Noise2music: Text-condit...

    2023 arXiv

  12. [20]

    Jairazbhoy

    N.A. Jairazbhoy. 1971. https://books.google.ae/books?id=0A0wAQAAIAAJ The R \=a gs of North Indian Music: Their Structure and Evolution . Wesleyan University Press

    1971

  13. [21]

    Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. https://doi.org/10.18653/v1/2020.acl-main.560 The state and fate of linguistic diversity and inclusion in the NLP world . In Proceedings of the 58th Annual Meeting of the Association for Co...

    2020 doi

  14. [22]

    Plumbley

    Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D. Plumbley. 2020. https://doi.org/10.1109/TASLP.2020.3030497 Panns: Large-scale pretrained audio neural networks for audio pattern recognition . IEEE Transactions on Audio, Speech, and Language Processing,...

    2020

  15. [23]

    Yun-Han Lan, Wen-Yi Hsiao, Hao-Chung Cheng, and Yi-Hsuan Yang. 2024. Musicongen: Rhythm and chord control for transformer-based text-to-music generation. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)

    2024

  16. [24]

    Peike Patrick Li, Boyu Chen, Yao Yao, Yikai Wang, Allen Wang, and Alex Wang. 2024. Jen-1: Text-guided universal music generation with omnidirectional diffusion models. In 2024 IEEE Conference on Artificial Intelligence (CAI), pages 762--769. IEEE

    2024

  17. [25]

    Shuyu Li and Yunsick Sung. 2021. Inco-gan: variable-length music generation method based on inception model-based conditional gan. Mathematics, 9(4):387

    2021

  18. [26]

    Liwei Lin, Gus Xia, Yixiao Zhang, and Junyan Jiang. 2024. https://doi.org/10.24963/ijcai.2024/851 Arrange, inpaint, and refine: Steerable long-term music audio generation and editing via content-based controls . In Proceedings of the Thirty-Third International Joint Conference...

    2024 doi

  19. [27]

    Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. 2023. AudioLDM : Text-to-audio generation with latent diffusion models. Proceedings of the International Conference on Machine Learning, pages 21450--21474

    2023

  20. [28]

    Ilya Loshchilov and Frank Hutter. 2017. https://api.semanticscholar.org/CorpusID:53592270 Decoupled weight decay regularization . In International Conference on Learning Representations

    2017

  21. [29]

    Holger Lund. 2019. https://doi.org/10.56513/QGAV8825-3 Decolonizing pop music . https://norient.com/

    2019 doi

  22. [30]

    Jan Melechovsky, Zixun Guo, Deepanway Ghosal, Navonil Majumder, Dorien Herremans, and Soujanya Poria. 2024. https://doi.org/10.18653/v1/2024.naacl-long.459 Mustango: Toward controllable text-to-music generation . In Proceedings of the 2024 Conference of the North American Chap...

    2024 doi

  23. [31]

    mrfakename, Vaibhav Srivastav, Clémentine Fourrier, Lucain Pouget, Yoach Lacombe, main, and Sanchit Gandhi. 2024. Text to speech arena. https://huggingface.co/spaces/TTS-AGI/TTS-Arena

    2024

  24. [32]

    Jonas Pfeiffer, Andreas R \"u ckl \'e , Clifton Poth, Aishwarya Kamath, Ivan Vuli \'c , Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020. https://doi.org/10.18653/v1/2020.emnlp-demos.7 A dapter H ub: A framework for adapting transformers . In Proceedings of the 2020 Co...

    2020 doi

  25. [33]

    Alastair Porter, Mohamed Sordo, and Xavier Serra. 2013. http://hdl.handle.net/10230/32251 Dunya: a system for browsing audio music collections exploiting cultural context

    2013

  26. [34]

    Alec Radford, Ilya Sutskever, et al. 2020. https://openai.com/research/jukebox Jukebox: A generative model for music . OpenAI Blog, 1(5)

    2020

  27. [35]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015, pages 234--241, Cham. Springer International Publishing

    2015

  28. [36]

    Flavio Schneider, Ojasv Kamal, Zhijing Jin, and Bernhard Sch \"o lkopf. 2024. Mo \^u sai: Efficient text-to-music diffusion models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8050--8068

    2024

  29. [37]

    Xavier Serra. 2014. Creating research corpora for the computational study of music: the case of the compmusic project. In AES 53rd International Conference: Semantic Audio; 2014 Jan 27-29; London, UK. New York: Audio Engineering Society; 2014. Article number 1-1 [9 p.]. Audio ...

    2014

  30. [38]

    K.L. Signell. 2008. https://books.google.ae/books?id=-G5MPgAACAAJ Makam: Modal Practice in Turkish Art Music . Usul Editions

    2008

  31. [39]

    Ajay Srinivasamurthy, Sankalp Gulati, Rafael Caro Repetto, and Xavier Serra. 2021. Saraga: open datasets for research on indian art music. Empirical Musicology Review, 16(1):85--98

    2021

  32. [40]

    Joseph P. Swain. 1995. https://doi.org/10.1093/mq/79.2.281 The Concept of Musical Syntax . The Musical Quarterly, 79(2):281--308

    1995 doi

  33. [41]

    Or Tal, Alon Ziv, Itai Gat, Felix Kreuk, and Yossi Adi. 2024. https://arxiv.org/abs/2406.10970 Joint audio and symbolic conditioning for temporally controlled text-to-music generation . Preprint, arXiv:2406.10970

    2024 arXiv

  34. [42]

    Shzr Ee Tan. 2021. https://doi.org/10.1080/17411912.2021.1938445 Special issue: Decolonising music and music studies . Ethnomusicology Forum, 30(1):4--8

    2021

  35. [43]

    Xiaodong Tan, Mathis Antony, and H Kong. 2020. Automated music generation for visual art through emotion. In ICCC, pages 247--250

    2020

  36. [44]

    Yan Tao, Olga Viberg, Ryan S Baker, and René F Kizilcec. 2024. https://doi.org/10.1093/pnasnexus/pgae346 Cultural bias and cultural alignment of large language models . PNAS Nexus, 3(9):pgae346

    2024 doi

  37. [45]

    Intercontinental Music Awards Team. 2023. https://www.intercontinentalmusicawards.com/music-and-ai-the-pros-cons-and-ethical-implications/ Music and ai: The pros, cons, and ethical implications . Intercontinental Music Awards

    2023

  38. [46]

    Yixiao Zhang, Yukara Ikemiya, Woosung Choi, Naoki Murata, Marco A Mart \' nez-Ram \' rez, Liwei Lin, Gus Xia, Wei-Hsiang Liao, Yuki Mitsufuji, and Simon Dixon. 2024. Instruct-musicgen: Unlocking text-to-music editing for music language models via instruction tuning. arXiv prep...

    2024 arXiv

  39. [47]

    Alon Ziv, Itai Gat, Gael Le Lan, Tal Remez, Felix Kreuk, Jade Copet, Alexandre D \'e fossez, Gabriel Synnaeve, and Yossi Adi. 2024. Masked audio generation using a single non-autoregressive transformer. In International Conference on Learning Representations (ICLR)

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.