Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Enhancing Lyrics Transcription on Music Mixtures with Consistency Loss

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper shows that adding a consistency loss between vocal and mixture encoder outputs during LoRA fine-tuning improves lyrics transcription on music mixtures, with L2 loss at weight 1.0 giving the best overall word error rate without…

desk verdict Solid empirical ALT paper with a modest, credible win for consistency loss; the train/test overlap question and missing significance tests are the real soft spots. read the letter →

arxiv 2506.02339 v1 pith:OUTBJNLZ submitted 2025-06-03 eess.AS cs.SD

Jiawen Huang , Felipe Sousa , Emir Demirel , Emmanouil Benetos , Igor Gadelha This is my paper
classification eess.AScs.SD
keywords automaticlyricstranscriptionlow-rankadaptationconsistencylossdual-domainfine-tuningWhispermusicmixturesworderrorratesingingvoiceseparation
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

Automatic lyrics transcription from full music mixes lags behind speech recognition because the accompaniment interferes with the singing voice, and most strong systems rely on source separation as a preprocessing step. This paper claims that a parameter-efficient LoRA fine-tune of the Whisper speech model, trained on both separated vocals and music mixtures, can transcribe mixtures directly. The key addition is a consistency loss that pushes the encoder's representations of a vocal-only input and the corresponding mixture toward each other. On the Multi-Lang Jamendo test set plus new Italian and Portuguese sets, the configuration with L2 consistency loss at weight 1.0 reaches the best overall-mixture word error rate, slightly beating the vocal-only model and clearly beating unadapted Whisper. If the result holds, better lyrics transcription on mixtures does not have to wait for a separate separation stage.

What carries the argument

The load-bearing mechanism is a dual-path fine-tuning loop. During training, the same Whisper encoder processes a separated vocal segment $X^v$ and the corresponding mixture $X^m$, producing representations $E^v$ and $E^m$; each path also computes its own cross-entropy transcription loss against the ground-truth lyrics. The paper adds a consistency loss $L_{\mathrm{CNS}} = l(E^v, E^m)$ with $l$ either L1 or L2, and trains on $L = (L^v_{\mathrm{ALT}} + L^m_{\mathrm{ALT}})/2 + w L_{\mathrm{CNS}}$. This term is what encourages the mixture-path encoder to suppress accompaniment information and produce vocal-like representations. LoRA adapters keep the update parameter-efficient, and decoding uses sliding 30-second windows in Whisper's long-form mode.

What would settle it

Check every track in the 118-song test set against the DALI v2.0 and MulJam v2.0 training track IDs, artist and song metadata, or audio fingerprints; if any test track appears in training, re-run the L2-loss w=1.0 and vocal-only models on the remaining disjoint subset and see whether the consistency-loss WER advantage persists.

Watch

Extended reading notes

Core claim

On its own terms, the paper reports that fine-tuning Whisper large-v2 with LoRA on singing voice improves lyrics transcription, and that how the vocal and mixture domains are combined matters. Naively averaging cross-entropy losses on paired vocal and mixture inputs underperforms vocal-only fine-tuning, but adding a consistency term $L_{\mathrm{CNS}}$ between the two encoder representations, with either L1 or L2 distance, consistently lowers word error rates on mixture inputs. The best configuration, L2 loss with weight $w=1.0$, reaches 32.36% WER on the overall mixture test set, beating the vocal-only model (32.85%) and unadapted Whisper (35.42%), and it does so without singing-voice separation at inference. The gains are described as modest, and consistency loss does not help the underrepresented Italian and Portuguese sets, where vocal-only training remains best.

Load-bearing premise

The reported mixture gains assume that the 79-song Multi-Lang Jamendo test set and the MulJam v2.0 training set share no overlapping songs sourced from MTG-Jamendo; if any overlap exists, the WER improvement could be inflated and the central claim would not generalize to unseen music.

Editorial extensions

If this is right

  • If the claim holds, music-mixture lyrics transcription can improve without a singing-voice separation front-end, removing a separately trained model and its errors from inference.
  • Since both L1 and L2 consistency variants beat the paired dual-domain baseline on English and overall mixture WER, the alignment objective itself, not the particular distance, carries the improvement.
  • The best consistency-loss model (L2, w=1.0) also holds vocal-transcription quality close to the vocal-only model, so the alignment does not sacrifice separated-vocal performance on the overall set.
  • The method's failure to improve Italian and Portuguese, which have less than 5% training share, indicates that low-resource languages remain better served by vocal-only fine-tuning.

Reading between the lines

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

  • A direct extension would test the same alignment with other ASR foundation models, such as Wav2Vec2 or WavLM, since the paper only reports Whisper large-v2 with LoRA rank 8.
  • If the encoder representations really converge, the mixture-path features could become useful for other vocal-centric tasks such as lyric-to-audio alignment or singer identification; the paper does not explore downstream uses.
  • The marginal margin over vocal-only raises the question whether a stronger alignment penalty, for example on decoder outputs or with contrastive negatives, would widen the gap; this is an editorial hypothesis, not a paper claim.
  • Given the shared MTG-Jamendo origin of training and test data, an independent evaluation on a verified disjoint dataset would clarify whether the reported gains are real or partly an artifact of song overlap; the paper does not perform that check.
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 / 5 minor

Summary. The paper proposes a parameter-efficient fine-tuning approach for automatic lyrics transcription (ALT) based on LoRA applied to Whisper large-v2. The method uses two input paths, separated vocals and music mixtures, and adds a consistency loss L_CNS = l(E_v, E_m) between the encoder representations of the two paths (Eq. 2), combined with the averaged path-wise cross-entropy losses (Eq. 3). Experiments on Multi-Lang Jamendo, plus newly annotated Italian and Portuguese evaluation sets, compare single-domain, dual-domain, and consistency-loss variants. The central claim, stated in Section 5.3, is that adding either an L1 or L2 consistency loss consistently improves WER on mixtures relative to the paired 'both' baseline, with L2 weight 1.0 achieving the best overall mix WER and outperforming the vocal-only model, all without requiring source separation at inference.

Significance. If the result holds, the contribution is a modest but useful step: adapting a general ASR foundation model to lyrics transcription with only LoRA, while avoiding singing-voice separation at inference. The paper also releases new annotations for two languages and a refined version of existing datasets, which is a concrete reproducibility asset. The consistency of the effect across all six consistency-loss configurations is encouraging. However, the empirical support is currently thin: there are no confidence intervals, significance tests, or multiple-seed runs, and an unverified train/test disjointness assumption underlies the evaluation. The central claim is plausible and within current consensus; the main need is validation.

major comments (3)
  1. [§4.1, Table 2] The paper never states that the Multi-Lang Jamendo test songs (79 tracks) and the MulJam v2.0 training songs (5,599 tracks) are disjoint, even though both are sourced from MTG-Jamendo. This is load-bearing for the Section 5.3 claim that consistency loss improves mixture WER on unseen music. The authors should check track-ID overlap between the released evaluation annotations and the MulJam training split and report the result; if any test songs appear in training, the reported WER gains (e.g., L2 w=1.0 overall mix 32.36 vs. 'both' 33.79) could partly reflect memorization rather than generalization.
  2. [§5.3, Table 3] All six consistency-loss configurations beat the 'both' baseline, but the reported differences are small (1-2 WER points) on a test set of 118 songs, and no confidence intervals, bootstrap tests, paired significance tests, or multiple-seed runs are provided. The claim that L2 w=1.0 'outperforms all other models, including voc only' therefore lacks statistical support. In addition, the consistency-loss weight w and loss function appear to have been selected on the same test set; the authors should either describe a validation-based selection procedure or explicitly acknowledge the selection when interpreting the results.
  3. [§4.2] The data-cleaning description says that 'we transcribed the lyrics with an internal Music.AI pretrained model for MulJam v2.0 and relabeled the languages accordingly.' This creates ambiguity about whether the training targets y used in Eq. (3) are the original human annotations or automatically generated transcriptions. If the targets are machine-generated, the models could be adapting to the internal ASR model's error patterns, which would change the interpretation of the WER gains. Please clarify the source of the training labels and, if automatic transcriptions were used, quantify their agreement with the original annotations.
minor comments (5)
  1. [Abstract] The rendering 'na¨ıve' should be 'naïve'; the LaTeX accent appears mis-rendered.
  2. [Table 3 caption] The caption refers to bold and underlined bold text, but the formatting is not visible in the submitted version; please make the conventions explicit or fix the table rendering.
  3. [§4.2] The sentence 'This results in a total of 38 languages' is confusing because Table 2 lists MulJam v2.0 as having six languages; clarify whether 38 refers to the relabeled training set after automatic transcription and language reassignment.
  4. [§2.2, Table 1] The Music.AI SVS model is internal and unreleased; one sentence on its architecture or a qualitative description would help reproducibility and interpretation of the SVS preprocessing baselines.
  5. [§4.3] Please state explicitly whether the learning-rate search was performed on a held-out validation split or on the test set, since the learning rate differs between the 'voc only' model and the other models.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the consistency-loss improvement is an externally measured WER comparison, not an outcome imposed by the loss or by self-citation.

full rationale

The paper is an empirical study rather than a formal derivation. The central claim, that adding the consistency loss improves mixture WER relative to the 'both' baseline, is supported by held-out evaluations on Multi-Lang Jamendo plus newly annotated Italian and Portuguese sets. The consistency loss (Eq. 2) is a training regularizer between encoder representations of paired vocal and mixture inputs, and the reported WERs are external metrics not fed back into the objective; no equation imposes the reported result by construction. The dual-domain baselines ('random' and 'both') are trained under the same protocol, so the comparison is not a fitted parameter renamed as a prediction. The self-citations present (notably MARBLE/MulJam v2.0, cited at [23], and Jam-ALT/LyricWhiz-type prior work cited at [13-14]) supply datasets and prior art, not a uniqueness theorem or an ansatz that forces the claimed outcome. The remaining concerns, such as unstated train/test overlap between MulJam v2.0 and Multi-Lang Jamendo and the parameter search on the learning rate, bear on generalization and soundness but are not circularity: they do not make the reported WER gains true by definition or by reduction to the paper's own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical study, so the ledger contains modeling and data assumptions rather than derived axioms. The consistency weight w is the main tuned quantity; LoRA rank, alpha, dropout, and learning rate are fixed hyperparameters. Two proprietary Music.AI components (SVS and alignment tool) are assumed to produce reliable vocal references and labels. No new physical entities are introduced.

free parameters (3)
  • consistency weight w = 0.1, 1.0, 10.0; best L2 w=1.0
    Tuned on the test set; no held-out validation is reported. The best configuration is selected by inspecting Table 3.
  • learning rate = 5e-7 for most models, 1e-6 for voc only
    Selected by parameter search over [1e-7, 5e-7, 1e-6] in Section 4.3.
  • LoRA rank, alpha, dropout = rank=8, alpha=8, dropout=0.5
    Fixed hyperparameters chosen without reported justification; they affect adapter capacity and regularization.
assumptions (4)
  • domain assumption Whisper large-v2 encoder features are informative for singing voice and can be steered by LoRA fine-tuning.
    Section 3 invokes Whisper's pretrained representations as the substrate for ALT; no proof that LoRA has sufficient capacity.
  • domain assumption Music.AI SVS separated vocals are a faithful clean reference for the vocal path.
    Section 4.2 generates vocal tracks with an internal SVS model; if the separation is imperfect, the consistency loss aligns mixture features to imperfect references.
  • domain assumption The training and evaluation tracks from MTG-Jamendo are disjoint.
    Section 4.1 lists MulJam v2.0 and Multi-Lang Jamendo, both sourced from MTG-Jamendo, without stating that overlapping songs were removed.
  • domain assumption The internal lyrics alignment and language relabeling tools produce accurate annotations.
    Section 4.2 relies on Music.AI's alignment tool and pretrained model to refine DALI and MulJam labels; errors propagate into training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Lyrics Transcription on Music Mixtures with Consistency Loss." pith.science (2026). https://pith.science/paper/OUTBJNLZ

@misc{pith2026250602339,
  author       = {Pith},
  title        = {Pith review of: Enhancing Lyrics Transcription on Music Mixtures with Consistency Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUTBJNLZ}},
  note         = {Machine review of arXiv:2506.02339}
}
read the original abstract

Automatic Lyrics Transcription (ALT) aims to recognize lyrics from singing voices, similar to Automatic Speech Recognition (ASR) for spoken language, but faces added complexity due to domain-specific properties of the singing voice. While foundation ASR models show robustness in various speech tasks, their performance degrades on singing voice, especially in the presence of musical accompaniment. This work focuses on this performance gap and explores Low-Rank Adaptation (LoRA) for ALT, investigating both single-domain and dual-domain fine-tuning strategies. We propose using a consistency loss to better align vocal and mixture encoder representations, improving transcription on mixture without relying on singing voice separation. Our results show that while na\"ive dual-domain fine-tuning underperforms, structured training with consistency loss yields modest but consistent gains, demonstrating the potential of adapting ASR foundation models for music.

Figures

Figures reproduced from arXiv: 2506.02339 by the authors.

Figure 1
Figure 1. Diagram of dual-domain fine-tuning with consistency loss. is required. However, full fine-tuning can be computation￾ally expensive and may be prone to overfitting, especially with smaller train datasets. Parameter-efficient fine-tuning (PEFT) techniques, such as Low-Rank Adaptation (LoRA) [9], address this challenge by updating only a small subset of model pa￾rameters while keeping the majority frozen. PEFT on found… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    Introduction Automatic Lyrics Transcription (ALT) is the task of recogniz- ing the linguistic content in singing voice recordings. It is the equivalent of automatic speech recognition (ASR) for speech, but presents additional challenges due to the unique character- istics of singing voice and the interference caused by music ac- companiment. Despite signi...

    work page Pith review arXiv 2025

  2. [2]

    PEFT on foundation models As ASR models get bigger, they require more and more com- putational resources which often may not be available for re- searchers

    Preliminaries 2.1. PEFT on foundation models As ASR models get bigger, they require more and more com- putational resources which often may not be available for re- searchers. To address the challenges of fine-tuning very large- scale transformers, parameter-efficient fine-tuning (PEFT) tech- niques have become popular. PEFT reduces the number of trainabl...

    work page

  3. [3]

    Method Due to their large-scale pretraining feature, the encoder repre- sentations of the open-source checkpoints for the Whisper mod- els are shown to capture rich background sound information, making them useful for tasks such as audio tagging [21]. When a singing voice is accompanied by an accompaniment, this ad- ditional musical information is also en...

    work page

  4. [4]

    Datasets For training, we use two datasets commonly used in recent ALT research, namely DALI v2.0 [22] and MulJam v2.0 [23]

    Experiments 4.1. Datasets For training, we use two datasets commonly used in recent ALT research, namely DALI v2.0 [22] and MulJam v2.0 [23]. DALI v2.0 is a multimodal, multilingual dataset for polyphonic mu- sic retrieved from YouTube, with time-aligned lyrics. To en- sure language diversity, we selected the five languages with at least 200 songs each (E...

    work page

  5. [5]

    Single-domain fine-tuning The top section in Table 3 presents the WERs for the base- line models

    Results and discussion 5.1. Single-domain fine-tuning The top section in Table 3 presents the WERs for the base- line models. We can make several observations. Both voc only and mix only outperform raw Whisper on overall mix- ture and overall vocals, indicating that fine-tuning on singing voice data is beneficial. However, for the English subset, per- for...

    work page

  6. [6]

    Conclusion In this work, we explored the use of LoRA finetuning on Whis- per for improving ALT and proposed a consistency loss for further improvements on music mixtures. Our results show that while na¨ıve dual-domain fine-tuning is ineffective, consis- tency loss provides modest but consistent improvements by bet- ter aligning vocal and mixture represent...

    work page

  7. [7]

    The authors would like to thank Vin´ıcius Marques, Angelo Maugeri, Ana Rachel Melo Nascimento, and Ana Lu ´ısa Runze from Music.AI for their help with data annotation and cleaning

    Acknowledgements JH is a research student at the UKRI Centre for Doctoral Train- ing in Artificial Intelligence and Music, supported jointly by UK Research and Innovation [grant number EP/S022694/1] and Queen Mary University of London. The authors would like to thank Vin´ıcius Marques, Angelo Maugeri, Ana Rachel Melo Nascimento, and Ana Lu ´ısa Runze from...

    work page

  8. [8]

    Automatic recognition of lyrics in singing,

    A. Mesaros and T. Virtanen, “Automatic recognition of lyrics in singing,” EURASIP J. Audio Speech Music. Process. , vol. 2010, 2010

    work page 2010

Show all 33 references
  1. [9]

    Mstre-net: Multistream- ing acoustic modeling for automatic lyrics transcription,

    E. Demirel, S. Ahlb ¨ack, and S. Dixon, “Mstre-net: Multistream- ing acoustic modeling for automatic lyrics transcription,” in Pro- ceedings of the 22nd International Society for Music Information Retrieval Conference, ISMIR 2021, Online, November 7-12, 2021, 2021, pp. 151–158

    2021

  2. [10]

    Polyscriber: Integrated fine-tuning of extractor and lyrics transcriber for polyphonic music,

    X. Gao, C. Gupta, and H. Li, “Polyscriber: Integrated fine-tuning of extractor and lyrics transcriber for polyphonic music,” IEEE ACM Trans. Audio Speech Lang. Process. , vol. 31, pp. 1968– 1981, 2023

    1968

  3. [11]

    Wavlm: Large- scale self-supervised pre-training for full stack speech process- ing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “Wavlm: Large- scale self-supervised pre-training for full stack speech process- ing,” IEEE J. Sel. Top. Si...

    2022

  4. [12]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” in Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Sys- tems 2020, NeurIPS 2020, Decembe...

    2020

  5. [13]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” in International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , ser. Pro- ceedings of Machine Learning...

    2023

  6. [14]

    Mir-mlpop: A mul- tilingual pop music dataset with time-aligned lyrics and audio,

    J. Wang, C. Wang, C. Leong, and J. R. Jang, “Mir-mlpop: A mul- tilingual pop music dataset with time-aligned lyrics and audio,” in IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2024, Seoul, Republic of Korea, April 14-19,

    2024

  7. [15]

    Hybrid transformers for music source separation,

    S. Rouard, F. Massa, and A. D ´efossez, “Hybrid transformers for music source separation,” in IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023. IEEE, 2023, pp. 1–5

    2023

  8. [16]

    Transfer learning of wav2vec 2.0 for automatic lyric transcription,

    L. Ou, X. Gu, and Y . Wang, “Transfer learning of wav2vec 2.0 for automatic lyric transcription,” inProceedings of the 23rd Interna- tional Society for Music Information Retrieval Conference, ISMIR 2022, Bengaluru, India, December 4-8, 2022, 2022, pp. 891–899

    2022

  9. [17]

    Lora: Low-rank adaptation of large lan- guage 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 lan- guage models,” in The Tenth International Conference on Learn- ing Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022

    2022

  10. [18]

    Parameter- efficient fine-tuning for large models: A comprehensive survey,

    Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, “Parameter- efficient fine-tuning for large models: A comprehensive survey,” Transactions on Machine Learning Research , 2024. [Online]. Available: https://openreview.net/forum?id=lIsCS8b6zj

    2024

  11. [19]

    Speech and language recognition with low-rank adaptation of pretrained models,

    A. Prasad, S. Madikeri, D. Khalil, P. Motlicek, and C. Schuep- bach, “Speech and language recognition with low-rank adaptation of pretrained models,” inInterspeech 2024, 2024, pp. 2825–2829

    2024

  12. [20]

    Enhancing two- stage finetuning for speech emotion recognition using adapters,

    Y . Gao, H. Shi, C. Chu, and T. Kawahara, “Enhancing two- stage finetuning for speech emotion recognition using adapters,” in ICASSP 2024 - 2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2024, pp. 11 316– 11 320

    2024

  13. [21]

    Lyricwhiz: Robust multilingual zero-shot lyrics transcription by whispering to chatgpt,

    L. Zhuo, R. Yuan, J. Pan, Y . Ma, Y . Li, G. Zhang, S. Liu, R. B. Dannenberg, J. Fu, C. Lin, E. Benetos, W. Chen, W. Xue, and Y . Guo, “Lyricwhiz: Robust multilingual zero-shot lyrics transcription by whispering to chatgpt,” in Proceedings of the 24th International Society for...

    2023 doi

  14. [22]

    Lyrics transcrip- tion for humans: A readability-aware benchmark,

    O. C ´ıfka, H. Schreiber, L. Miner, and F. St¨oter, “Lyrics transcrip- tion for humans: A readability-aware benchmark,” in Proceed- ings of the 25th International Society for Music Information Re- trieval Conference, ISMIR 2024, San Francisco, California, USA and Online, Novem...

    2024

  15. [23]

    MARBLE: music audio representation benchmark for universal evaluation,

    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. Bene- tos, A. Ragni, N. Gyenge, R. B. Dannenberg, W. Chen, G. Xia, W. Xue, S. Liu, S. Wang, R. Liu, Y . Guo, and J. Fu, “MARBLE: music audio representation ben...

    2023

  16. [24]

    Contrastive learning-based audio to lyrics alignment for multiple languages,

    S. Durand, D. Stoller, and S. Ewert, “Contrastive learning-based audio to lyrics alignment for multiple languages,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  17. [25]

    Dual-path adaptation of pretrained fea- ture extraction module for robust automatic speech recognition,

    H. Shi and T. Kawahara, “Dual-path adaptation of pretrained fea- ture extraction module for robust automatic speech recognition,” in Proc. Interspeech 2024, 2024, pp. 2850–2854

    2024

  18. [26]

    A noise-robust self-supervised pre-training model based speech rep- resentation learning for automatic speech recognition,

    Q. Zhu, J. Zhang, Z. Zhang, M. Wu, X. Fang, and L. Dai, “A noise-robust self-supervised pre-training model based speech rep- resentation learning for automatic speech recognition,” in IEEE International Conference on Acoustics, Speech and Signal Pro- cessing, ICASSP 2022, Virt...

    2022

  19. [27]

    Dual-path style learn- ing for end-to-end noise-robust speech recognition,

    Y . Hu, N. Hou, C. Chen, and E. S. Chng, “Dual-path style learn- ing for end-to-end noise-robust speech recognition,” in 24th An- nual Conference of the International Speech Communication As- sociation, Interspeech 2023, Dublin, Ireland, August 20-24, 2023. ISCA, 2023, pp. 2918–2922

    2023

  20. [28]

    CR-CTC: Consistency regularization on CTC for improved speech recognition,

    Z. Yao, W. Kang, X. Yang, F. Kuang, L. Guo, H. Zhu, Z. Jin, Z. Li, L. Lin, and D. Povey, “CR-CTC: Consistency regularization on CTC for improved speech recognition,” inThe Thirteenth Interna- tional Conference on Learning Representations , 2025. [Online]. Available: https://op...

    2025

  21. [29]

    Whisper-at: Noise-robust automatic speech recognizers are also strong gen- eral audio event taggers,

    Y . Gong, S. Khurana, L. Karlinsky, and J. R. Glass, “Whisper-at: Noise-robust automatic speech recognizers are also strong gen- eral audio event taggers,” in 24th Annual Conference of the Inter- national Speech Communication Association, Interspeech 2023, Dublin, Ireland, Aug...

    2023

  22. [30]

    Creating DALI, a large dataset of synchronized audio, lyrics, and notes,

    G. Meseguer-Brocal, A. Cohen-Hadria, and G. Peeters, “Creating DALI, a large dataset of synchronized audio, lyrics, and notes,” Transactions of the International Society for Music Information Retrieval, vol. 3, no. 1, 2020

    2020

  23. [32]

    The mtg-jamendo dataset for automatic music tagging,

    D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra, “The mtg-jamendo dataset for automatic music tagging,” in Ma- chine Learning for Music Discovery Workshop, International Conference on Machine Learning (ICML 2019), Long Beach, CA, United States, 2019

    2019

  24. [33]

    Patel, Music, Language, and the Brain

    A. Patel, Music, Language, and the Brain . Oxford University Press, USA, 2010. [Online]. Available: https://books.google.co. uk/books?id=qekVDAAAQBAJ

    2010

  25. [2024]

    1366–1370

    IEEE, 2024, pp. 1366–1370

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.