Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Improving Audio Event Recognition with Consistency Regularization

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper argues that consistency regularization (CR) — a label-free loss that asks a model to give the same predictions for two augmented copies of the same audio — brings reliable gains to multi-label audio event recognition. On AudioSet,

desk verdict Plausible, well-ablated application of consistency regularization to audio event recognition, but single-run numbers and a text/table mismatch mean the gains aren't proven yet. read the letter →

arxiv 2509.10391 v1 pith:CVYM5EFM submitted 2025-09-12 cs.SD cs.AI

Shanmuka Sadhu , Weiran Wang This is my paper
classification cs.SDcs.AI
keywords consistencyregularizationaudioeventrecognitionSetdataaugmentationsemi-supervisedlearningmeanaverageprecisionmulti-labelclassificationtransformerfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that consistency regularization (CR) — a label-free loss that asks a model to give the same predictions for two augmented copies of the same audio — brings reliable gains to multi-label audio event recognition. On AudioSet, it lifts mean average precision from 37.9 to 39.6 on the 20k training subset and from 44.7 to 46.9 on the near-2M subset, both over baselines that already use heavy data augmentation. It also shows a further gain to 40.1 mAP in a semi-supervised setting that treats the large set as unlabeled, and reports improvements even without pretrained initialization. A sympathetic reader would take this as evidence that enforcing prediction invariance to augmentations is a broadly useful training signal for audio.

What carries the argument

The consistency regularization loss itself: given two augmented views x1 and x2 of the same input, the model produces probabilities p1 and p2, and the loss computes 0.5*(sg(p1) log p2 + (1-sg(p1)) log(1-p2)) plus the symmetric term, where sg denotes stop-gradient. For k > 2 views, all unordered pairs contribute equally. This loss is added to the standard binary cross-entropy loss, and because it is label-free it can be applied to unlabeled audio in semi-supervised training. The paper also tunes the input augmentations (mixup, SpecAugment, random erasing) and the number of views.

What would settle it

Re-run the AS-20k and AS-2M fine-tuning with and without CR over at least five seeds and compare the distributions of mAP. If the mean difference is smaller than the seed-to-seed standard deviation, the claimed consistent improvement is not distinguishable from training noise.

Watch

Extended reading notes

Core claim

The central claim is that CR is a generally effective training regularizer for audio event recognition. The paper defines CR as a symmetric cross-entropy between the prediction distributions of two augmented views, with a stop-gradient operator to prevent degenerate solutions, and generalizes it to any number of views by averaging pairwise terms. Combined with the binary cross-entropy supervised loss, this term consistently improves results on AudioSet across training set sizes (20k and 1.8M), across augmentations, and with and without pretrained model weights. Because the CR loss does not require labels, it also plugs into a semi-supervised pipeline where only 20k samples are labeled and th

Load-bearing premise

The central 'consistent improvement' claim rests on single-run mAP numbers; if run-to-run variance on AudioSet fine-tuning is comparable to the reported 1.7–2.1 mAP gaps, the advantage could vanish.

Editorial extensions

If this is right

  • On AudioSet, CR improves mAP from 37.9 to 39.6 (AS-20k) and from 44.7 to 46.9 (AS-2M) over already-strong augmented baselines.
  • The semi-supervised variant adds another 0.4 mAP on AS-20k by exploiting 1.8M unlabeled clips.
  • The gain does not depend on pretrained weights: without pretraining, CR adds 2.1 mAP on AS-20k and 2.6 mAP on AS-2M.
  • More CR views help on the small training set (2 vs 6 augmentations), while two views suffice on the large set.
  • Random erasing, an image augmentation, helps the small-set CR setup but not the large-set one.

Reading between the lines

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

  • If the effect is real, CR could be combined with other label-free objectives such as masked prediction or contrastive learning to push audio representation quality further, since it occupies a different niche: invariance at the probability level rather than at the embedding level.
  • The technique should transfer to other multi-label sequence tasks (e.g., sound event detection with localization, or polyphonic music tagging), where the same 'same input, same label distribution' logic applies.
  • A natural testable extension: vary augmentation strength during CR training and measure whether the optimal strength tracks the label-noise level of the dataset, which would indicate CR is mainly suppressing overfitting to augmentation artifacts.
  • Because the reported numbers are single runs, an independent multi-seed replication would be the quickest way to separate a real effect from training noise.
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. This paper proposes adding consistency regularization (CR) to the AudioMAE model for audio event recognition. The CR loss encourages agreement between model predictions on two or more augmented views of the same spectrogram; the supervised total loss is L_BCE + λ L_CR. The authors ablate λ, mixup ratio, random erasing probability, and number of augmentations on AudioSet-20k and on a 1.78M-sample AudioSet-2M setup, and extend CR to a semi-supervised setting with 20k labeled samples and the large set treated as unlabeled. They report test mAP improvements from 37.9 to 39.6 on AS-20k and 44.7 to 46.9 on AS-2M over a reproduced AudioMAE baseline, with 40.1 mAP for the semi-supervised model.

Significance. If the reported improvements are reproducible, this is a practically useful contribution: a simple plug-in auxiliary loss that improves an already augmentation-heavy ViT baseline by 1.7–2.1 mAP, and a principled way to exploit unlabeled AudioSet data. The paper is honest that its reproduced large-set baseline uses fewer samples than the published AudioMAE (1.78M vs. 2M), and it releases code. I saw no circular reasoning: the CR loss is not derived from the evaluation metric, and the final test numbers are from a held-out test set. However, the central claims currently rest on single-run evaluations and on a loss definition that is mis-specified as written, so the strength of the evidence is insufficient.

major comments (3)
  1. [§3.2, Eqs. (2)–(3)] The CR loss is mis-specified as written. A binary cross-entropy term should be −[q log p + (1−q) log(1−p)]. Equation (2) is written as sg(p1) log p2 − (1−sg(p1)) log(1−p2); the missing overall sign and the minus before the second term make the loss negative and drive p2 away from p1 when minimized. Unless this is a typesetting error and the implementation uses the standard BCE, Eq. (4) is not a consistency loss and the reported results cannot be attributed to it. Please correct the formula and clarify its relationship to the KL/BCE losses used in prior CR work.
  2. [§4.4, Table 3(a)] The text states that 'the most optimal coefficient setup is 1.5 for λ1 and 1.0 for λ2', but Table 3(a) shows that this row gives 36.3 mAP, while λ1=1.0, λ2=1.5 gives 36.6 mAP. The claimed 0.4 mAP semi-supervised gain corresponds to the latter row, not the one named in the text. This internal contradiction must be resolved; as written, the reported optimal hyperparameters are ambiguous and the semi-supervised claim is not reproducible from the table.
  3. [§4.5, Table 4; overall claim] All reported improvements are single-run mAP numbers on one AudioSet split, with no error bars, multiple seeds, or significance tests. The headline gains (1.7 mAP on AS-20k, 2.1 mAP on AS-2M) are within the range of run-to-run variation that is typical when fine-tuning ViT-B on AudioSet. The paper's central claim of 'consistent improvement' therefore lacks direct statistical support. Moreover, hyperparameters (λ, mixup ratio, erasing probability, number of augmentations) are tuned on the same dev set across many ablation rows (Tables 1–3) without correction for multiple comparisons, which can inflate the final test difference. Please add repeated independent runs (at least 3 seeds) with mean±std for baseline vs. proposed method, or a paired test on the dev set, to support the claim of consistency.
minor comments (5)
  1. [§4.1] The sentence 'we had less training data (∼200K for AS-2M)' is unclear. State that the AS-2M training set here contains 1,783,977 samples, about 0.26M fewer than the original 2,042,985, and that this explains why the reproduced baseline (44.7 mAP) is lower than the published AudioMAE value (47.3 mAP).
  2. [§4.2, Table 1] Clarify whether the λ=0 row in Table 1(a) uses the same augmentation pipeline (mixup, random erasing, 556 labels) as the final baseline in Table 4. The 34.7 dev mAP vs. 37.9 test mAP gap makes the ablation and final setups not directly comparable.
  3. [§4.4] The sentence 'in 60 epochs we will see the 1.8M unsupervised set around 3 times' should be stated as approximately 2.7 times: 4×20k unlabeled samples per epoch × 60 epochs = 4.8M samples.
  4. [Throughout] Typos and wording: 'perfomring' (§3.3), 'comparision' (§4.2), 'an semi-supervise setup' (§4.5).
  5. [§2] When citing CR-CTC, specify that it uses frame-level KL divergence, while the proposed loss is a stop-gradient cross-entropy; the distinction matters because KL divergence is asymmetric and the loss in Eq. (4) is symmetric by construction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CR is an empirical loss applied to AudioSet, hyperparameters tuned on held-out dev data, and test mAP compared against reproduced baselines.

full rationale

The paper's central claim—that adding consistency regularization (CR) to AudioMAE fine-tuning improves AudioSet mAP—is an empirical result, not a derivation. The CR loss in Eqs. (2)–(5) is defined directly from model predictions on augmented views and contains no term depending on ground-truth test labels or the mAP evaluation metric. Hyperparameters (λ, mixup ratio, random-erasing probability, number of augmentations) are tuned on a held-out development set, as stated in Section 4.1: 'Our development set consists of 18,610 samples and is separate from our training sets.' The final test numbers in Table 4 are compared against a reproduced AudioMAE baseline with the same architecture, pretrained checkpoint, and label set (556 training classes, 527 evaluation classes). No parameter is fitted to the test labels, and the reported gains are not forced by the definition of the loss. The only external inputs are the AudioMAE checkpoint [12] and the CR idea from CR-CTC [1], neither authored by the present authors, so no load-bearing self-citation chain exists. There is an internal inconsistency between Section 4.4 text ('We find the most optimal coefficient setup to be 1.5 for λ1 and 1.0 for λ2') and Table 3(a), which shows λ1=1, λ2=1.5 as better (36.6 vs 36.3); this is a reporting/support concern, not circularity. The absence of multiple seeds is a statistical robustness issue, not a circularity issue. The derivation is self-contained with respect to the defined loss and evaluation protocol.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The paper contributes an empirical recipe; its free parameters are the hyperparameters tuned on the AudioSet dev set. No new theoretical entities are introduced.

free parameters (8)
  • CR coefficient lambda (AS-20k) = 2.0
    Tuned on dev set (Table 1a); best mAP 35.8 vs 34.7 baseline.
  • CR coefficient lambda (AS-2M) = 1.5
    Tuned on dev set (Table 2a); best mAP 46.6.
  • Semi-supervised lambda1 = 1.5 (text) / 1.0 (table best)
    Text states 1.5, but Table 3a shows best at lambda1=1.0, lambda2=1.5 (36.6 mAP) vs 36.3 at (1.5,1.0).
  • Semi-supervised lambda2 = 1.0 (text) / 1.5 (table best)
    Same inconsistency as lambda1.
  • Mixup ratio mu = 0.5
    Tuned on dev set (Table 1b); best 35.8.
  • Random erasing probability (AS-20k) = 0.25
    Tuned on dev set (Table 1c); best 36.0.
  • Number of augmentations (AS-20k) = 6
    Tuned on dev set (Table 1d); best 36.2.
  • Number of augmentations (AS-2M) = 2
    Tuned on dev set (Table 2c); more augments degraded performance.
assumptions (4)
  • domain assumption AudioSet dev set mAP is a reliable proxy for test performance and for model selection.
    The paper tunes hyperparameters on the dev set and reports only the chosen configuration's test mAP.
  • domain assumption The AudioMAE architecture and pretrained checkpoint are appropriate for AER.
    The paper builds entirely on AudioMAE and its checkpoint, assuming they are suitable.
  • domain assumption The CR loss with stop-gradient and BCE-based pseudo-labels provides a stable and meaningful training signal.
    The method takes CR-CTC's loss design and applies it without theoretical justification beyond the intuition of invariance.
  • domain assumption The augmentations (mixup, SpecAugment, random erasing) preserve the audio event semantics needed for CR.
    The paper empirically selects augmentations but does not prove they maintain the relevant information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Audio Event Recognition with Consistency Regularization." pith.science (2026). https://pith.science/paper/CVYM5EFM

@misc{pith2026250910391,
  author       = {Pith},
  title        = {Pith review of: Improving Audio Event Recognition with Consistency Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CVYM5EFM}},
  note         = {Machine review of arXiv:2509.10391}
}
abstract

Consistency regularization (CR), which enforces agreement between model predictions on augmented views, has found recent benefits in automatic speech recognition [1]. In this paper, we propose the use of consistency regularization for audio event recognition, and demonstrate its effectiveness on AudioSet. With extensive ablation studies for both small ($\sim$20k) and large ($\sim$1.8M) supervised training sets, we show that CR brings consistent improvement over supervised baselines which already heavily utilize data augmentation, and CR using stronger augmentation and multiple augmentations leads to additional gain for the small training set. Furthermore, we extend the use of CR into the semi-supervised setup with 20K labeled samples and 1.8M unlabeled samples, and obtain performance improvement over our best model trained on the small set.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 8 linked inside Pith

  1. [1]

    Many se- tups, including supervised and self-supervised audio representation learning, have shown significant results on AER datasets

    INTRODUCTION Audio Event Recognition (AER) plays a major role in real-world systems, such as wearable devices [2], smart home devices [3], surveillance systems [4–6], and human activity [7, 8]. Many se- tups, including supervised and self-supervised audio representation learning, have shown significant results on AER datasets. Early success used non pretr...

  2. [2]

    Supervised methods such as RDrop [22] and cosub [23] have demonstrated the advantage of CR by encouraging agreement between model predictions on different augmentations

    RELATED WORKS Consistency regularization (CR):Consistency Regularization has been useful in self-supervised methods, such as SimCLR [18], BYOL [19], MoCo [20] and SimSiam [21], which use different augmented views to learn semantic representations of unlabeled samples. Supervised methods such as RDrop [22] and cosub [23] have demonstrated the advantage of ...

    arXiv 2025

  3. [3]

    pseudo-label

    METHODS 3.1. Audio Event Recognition Audio Event Recognition (AER) is inherently a multi-label problem, i.e., each audio recording can be associated with multiple classes. We use binary cross entropy (BCE) loss for each class to predict its presence. For a training set comprisingNsamples andCevents, the BCE loss is defined as LBCE =− 1 N NX i=1 CX c=1 [yi...

  4. [4]

    Experimental Setup AudioSet:AudioSet [33] is a multi-label audio dataset with 2 mil- lion audio clips of length 10 seconds, totalling to 5,800 hours

    EXPERIMENTS 4.1. Experimental Setup AudioSet:AudioSet [33] is a multi-label audio dataset with 2 mil- lion audio clips of length 10 seconds, totalling to 5,800 hours. There are two commonly used supervised training sets: AS-20k is class- wise balanced and originally contained 22,176 audio samples [30], and AS-2M is unbalanced and originally had 2,042,985 ...

  5. [5]

    Additionally, we apply consistency regularization to a semi-supervised setting

    CONCLUSION & FUTURE WORK In this work, we introduce consistency regularization (CR) into au- dio event recognition through extensive experiments with AS-20k and AS-2M. Additionally, we apply consistency regularization to a semi-supervised setting. Our findings concluded that 1) consistency regularization in supervised and semi-supervised settings demon- s...

  6. [6]

    Cr-ctc: Consistency regularization on ctc for improved speech recognition,

    Zengwei Yao, Wei Kang, Xiaoyu Yang, Fangjun Kuang, Liy- ong Guo, Han Zhu, Zengrui Jin, Zhaoqing Li, Long Lin, and Daniel Povey, “Cr-ctc: Consistency regularization on ctc for improved speech recognition,” inICLR, 2025

    2025

  7. [7]

    Spatial audio processing with large language model on wear- able devices,

    A. Mishra, Y . Bai, P. Narayanasamy, N. Garg, and N. Roy, “Spatial audio processing with large language model on wear- able devices,”arXiv preprint arXiv:2504.08907, 2025

    arXiv 2025

  8. [8]

    Audio-based event recognition system for smart homes,

    A. Vafeiadis, K. V otis, D. Giakoumis, D. Tzovaras, L. Chen, and R. Hamzaoui, “Audio-based event recognition system for smart homes,” inProc. IEEE SmartWorld, 2017

    2017

Show all 41 references
  1. [9]

    homesound: Real-time audio event detection based on high performance computing for behaviour and surveillance remote monitoring,

    R. M. Alsina-Pag `es, J. Navarro, F. Al ´ıas, and M. Herv ´as, “homesound: Real-time audio event detection based on high performance computing for behaviour and surveillance remote monitoring,”Sensors, vol. 17, no. 4, 2017

    2017

  2. [10]

    Audio-assisted smart home security monitor- ing with few samples,

    H. Chi et al., “Audio-assisted smart home security monitor- ing with few samples,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2024

    2024

  3. [11]

    Smart home security through real-time audio event detection systems using convo- lutional neural network (cnn),

    Mark Daniel Dacer and Junrie Matias, “Smart home security through real-time audio event detection systems using convo- lutional neural network (cnn),” inProc. IEEE ICERA, 2025

    2025

  4. [12]

    Real-time sound event classification for human activity of daily living using deep neural network,

    A. H. Yuh and S. J. Kang, “Real-time sound event classification for human activity of daily living using deep neural network,” inProc. IEEE Int. Conf. Internet Things (iThings), 2021

    2021

  5. [13]

    Occupant behavior monitoring and emergency event detection in single-person households using deep learning-based sound recognition,

    Minhyuk Jung Seokho Chi, Jinwoo Kim, and Kyungjun Min, “Occupant behavior monitoring and emergency event detection in single-person households using deep learning-based sound recognition,”Building and Environment, 2020

    2020

  6. [14]

    Cnn architectures for large-scale audio clas- sification,

    S. Hershey et al., “Cnn architectures for large-scale audio clas- sification,” inICASSP, 2017

    2017

  7. [15]

    Panns: Large-scale pretrained audio neural net- works for audio pattern recognition,

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “Panns: Large-scale pretrained audio neural net- works for audio pattern recognition,”TASLP, 2020

    2020

  8. [16]

    Contrastive audio-visual masked autoencoder,

    Y . Gong et al., “Contrastive audio-visual masked autoencoder,” inICLR, 2023

    2023

  9. [17]

    Masked autoencoders that listen,

    Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feicht- enhofer, “Masked autoencoders that listen,” inNeurIPS, 2022

    2022

  10. [18]

    Clap learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “Clap learning audio concepts from natural language supervision,” in ICASSP, 2023

    2023

  11. [19]

    Ast: audio spectrogram transformer,

    Y . Gong, Y . Chung, and J. R. Glass, “Ast: audio spectrogram transformer,” inInterspeech, 2021

    2021

  12. [20]

    Equiav: Leveraging equivariance for audio-visual contrastive learning,

    J. Kim, H. Lee, K. Rho, J. Kim, and J. S. Chung, “Equiav: Leveraging equivariance for audio-visual contrastive learning,” arXiv preprint arXiv:2403.09502, 2024

    2024 arXiv

  13. [21]

    Lis- ten, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. Glass, “Lis- ten, think, and understand,”arXiv preprint arXiv:2305.10790, 2023

    2023 arXiv

  14. [22]

    Gama: A large audio-language model with ad- vanced audio understanding and complex reasoning abilities,

    S. Ghosh, S. Kumar, A. Seth, C. K. R. Evuru, U. Tyagi, S. Sak- shi, et al., “Gama: A large audio-language model with ad- vanced audio understanding and complex reasoning abilities,” arXiv preprint arXiv:2406.11768, 2024

    2024 arXiv

  15. [23]

    A simple framework for contrastive learning of visual representations,

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geof- frey Hinton, “A simple framework for contrastive learning of visual representations,” inICML, 2020

    2020

  16. [24]

    Bootstrap your own latent-a new approach to self-supervised learning,

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al., “Bootstrap your own latent-a new approach to self-supervised learning,” inNeurIPS, 2020

    2020

  17. [25]

    Momentum contrast for unsupervised visual repre- sentation learning,

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick, “Momentum contrast for unsupervised visual repre- sentation learning,” inCVPR, 2020

    2020

  18. [26]

    Exploring simple siamese rep- resentation learning,

    Xinlei Chen and Kaiming He, “Exploring simple siamese rep- resentation learning,” inCVPR, 2021

    2021

  19. [27]

    R-drop: Regularized dropout for neural networks,

    Lijun Wu, Juntao Li, Yue Wang, Qi Meng, Tao Qin, Wei Chen, Min Zhang, and Tie-Yan Liu, “R-drop: Regularized dropout for neural networks,” inNeurIPS, 2021

    2021

  20. [28]

    Co-training 2l submodels for visual recognition,

    Hugo Touvron, Matthieu Cord, Maxime Oquab, Piotr Bo- janowski, Jakob Verbeek, and Herv ´e J ´egou, “Co-training 2l submodels for visual recognition,” inCVPR, 2023

    2023

  21. [29]

    Deep convolutional neural net- works and data augmentation for environmental sound classi- fication,

    J. Salamon and J. P. Bello, “Deep convolutional neural net- works and data augmentation for environmental sound classi- fication,”IEEE Signal Processing Letters, vol. 24, no. 3, 2017

    2017

  22. [30]

    Random erasing data augmentation,

    Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang, “Random erasing data augmentation,” inAAAI, 2020

    2020

  23. [31]

    V ocal tract length perturbation (vtlp) improves speech recognition,

    Navdeep Jaitly and Geoffrey E. Hinton, “V ocal tract length perturbation (vtlp) improves speech recognition,” inICML, 2013

    2013

  24. [32]

    Audio aug- mentation for speech recognition,

    T. Ko, V . Peddinti, D. Povey, and S. Khudanpur, “Audio aug- mentation for speech recognition,” inINTERSPEECH, 2015

    2015

  25. [33]

    mixup: Beyond empirical risk minimization,

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz, “mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  26. [34]

    Bert: Pre- training of deep bidirectional transformers for language under- standing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre- training of deep bidirectional transformers for language under- standing,”arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  27. [35]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” inCVPR, 2022

    2022

  28. [36]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,” TASLP, 2021

    2021

  29. [37]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” inINTERSPEECH, 2019

    2019

  30. [38]

    Audio set: An ontology and human- labeled dataset for audio events,

    J. F. Gemmeke et al., “Audio set: An ontology and human- labeled dataset for audio events,” inICASSP, 2017

    2017

  31. [39]

    Ss- lam: Enhancing self-supervised models with audio mixtures for polyphonic soundscapes,

    T. Alex, S. Atito, A. Mustafa, M. Awais, and P. J. Jackson, “Ss- lam: Enhancing self-supervised models with audio mixtures for polyphonic soundscapes,” inICLR, 2025

    2025

  32. [40]

    Salmonn: Towards generic hearing abilities for large language models.,

    Guangzhi Sun Xianzhao Chen Tian Tan Wei Li Lu Lu Ze- jun MA Changli Tang, Wenyi Yu and Chao Zhang, “Salmonn: Towards generic hearing abilities for large language models.,” inICML, 2024

    2024

  33. [41]

    Qwen- audio: Advancing universal audio understanding via uni- fied large-scale audiolanguage models.,

    Xiaohuan Zhou Qian Yang Shiliang Zhang Zhijie Yan Chang Zhou Yunfei Chu, Jin Xu and Jingren Zhou, “Qwen- audio: Advancing universal audio understanding via uni- fied large-scale audiolanguage models.,”arXiv preprint arXiv:2311.07919., 2023

    2023 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.