The reviewed record of science sign in
Pith

arxiv: 2606.19203 · v1 · pith:J2MFAWQU · submitted 2026-06-17 · eess.AS

DASH: Dual-View Self-Distillation with Multi-Layer Hidden Representations for Robust Speech Recognition

Jaeeun Baik , Ui-Hyeop Shin , Jiwoon Lee , Woocheol Jeong , Hyung-Min Park This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-26 19:10 UTCgrok-4.3pith:J2MFAWQUrecord.jsonopen to challenge →

classification eess.AS
keywords automatic speech recognitionnoise robustnessself-distillationhidden representationsdual-view learningprototype assignment
0
0 comments X

The pith

Dual-view self-distillation on clean and noisy speech pairs improves recognition in noise while preserving clean accuracy.

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

The paper proposes a label-free pre-training approach that teaches a speech recognition model to treat clean and noisy versions of the same utterance as equivalent. It does so by distilling representations across multiple encoder layers and aligning the distributions of how those representations are assigned to prototypes. This targets the common problem where noise-robust training hurts performance on clean speech or overfits to particular noise types. The added stage requires only about four percent of the time of standard fine-tuning. If the approach holds, models could be made more reliable for real-world use without the usual accuracy trade-offs.

Core claim

DASH performs self-distillation by passing paired clean and noisy views through the same encoder, extracting hidden representations from multiple layers to span low-level acoustics to high-level semantics, and minimizing KL divergence between the prototype assignment distributions of the two views; this consistency learning occurs in a label-free pre-training stage before standard fine-tuning and yields improved noisy recognition without loss of clean accuracy.

What carries the argument

Dual-view self-distillation with multi-layer hidden representations, which enforces clean-noisy consistency by distilling across encoder layers and stabilizing training via KL divergence on prototype assignments.

If this is right

  • Recognition accuracy rises under diverse noisy conditions.
  • Clean-speech accuracy stays at the level achieved by standard fine-tuning.
  • The pre-training stage adds only a small fraction of the fine-tuning compute cost.
  • The model avoids overfitting to the specific corruptions used during training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same dual-view consistency idea could be tested on other sequence tasks that involve paired clean and degraded inputs.
  • Using multiple layers for distillation might reduce the need for task-specific data augmentation strategies.
  • If the prototype alignment step proves stable, it could replace some forms of supervised noise injection in production pipelines.

Load-bearing premise

Aligning multi-layer hidden representations and prototype distributions between clean and noisy views will produce features that generalize across noise types without creating new trade-offs or overfitting.

What would settle it

Apply the method to speech data containing a previously unseen noise type and measure whether clean-speech word error rate rises or noisy-speech word error rate fails to improve relative to a standard fine-tuned baseline.

Figures

Figures reproduced from arXiv: 2606.19203 by Hyung-Min Park, Jaeeun Baik, Jiwoon Lee, Ui-Hyeop Shin, Woocheol Jeong.

Figure 1
Figure 1. Figure 1: Overview of the DASH pipeline. invariant speech representations by comparing paired clean and noisy views of the same utterance. To achieve this, we instan￾tiate a dual-branch encoder architecture consisting of a teacher (clean) network and a student (noisy) network. The clean net￾work processes the unperturbed view xclean, extracting clean acoustic features, while the noisy network processes the aug￾mente… view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualization of encoder representations ex￾tracted from Layers 6 and 17 for the fine-tuning-only baseline (left) and DASH (right). Points are colored by utterance. solely to the final encoder layer. Under this constrained setting, the model’s generalization capability on challenging acous￾tic conditions significantly deteriorated, even falling behind the standard fine-tuning baseline. These results … view at source ↗
read the original abstract

Automatic Speech Recognition (ASR) often degrades in real-world noisy environments, making noise robustness essential for deployment. Supervised noise-augmented fine-tuning is a common remedy, but it can introduce a robustness-clean trade-off and overfit to specific corruptions, degrading recognition in clean conditions. We propose DASH, a self-distillation framework that improves robustness by learning clean--noisy consistency from paired views. DASH distills hidden representations from multiple encoder layers to capture features from low-level acoustics to high-level semantics, and stabilizes training by minimizing KL divergence between prototype assignment distributions of clean and noisy views. Experiments on LibriSpeech show that DASH consistently improves recognition under diverse noisy conditions while preserving clean accuracy, achieved by a label-free pre-training stage with minimal additional overhead (about 4% of fine-tuning time) beyond standard fine-tuning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes DASH, a self-distillation framework for robust automatic speech recognition. It performs label-free pre-training by distilling multi-layer hidden representations from paired clean and noisy views of the input, stabilized via KL divergence minimization between prototype assignment distributions of the two views. The central claim is that this yields consistent improvements in recognition accuracy under diverse noisy conditions on LibriSpeech while preserving clean-condition performance, at a cost of only ~4% additional compute relative to standard fine-tuning.

Significance. If the empirical results hold with the required controls, the contribution would be significant for practical ASR deployment: it targets the well-known robustness-clean accuracy trade-off without relying on supervised noise augmentation. The multi-layer distillation strategy (low-level acoustics to high-level semantics) and the use of prototype KL for training stability are technically interesting and could generalize to other sequence models. The reported low overhead is a concrete practical advantage.

major comments (2)
  1. [Abstract] The abstract asserts that DASH 'consistently improves recognition under diverse noisy conditions' yet supplies no quantitative results, baselines, error bars, or experimental protocol. This absence is load-bearing for the central claim, as the soundness assessment notes the lack of evidence to substantiate the no-trade-off and generalization assertions.
  2. [Method / Experiments] The pre-training procedure relies on clean-noisy view pairs, but the manuscript does not state whether the noise types (or noise distribution) used during the label-free pre-training stage are disjoint from those appearing in the evaluation conditions. Overlap would allow the multi-layer distillation plus KL objective to encode corruption-specific statistics rather than invariant acoustic-to-semantic mappings, directly threatening the generalization claim to 'unseen' diverse noises.
minor comments (2)
  1. [Method] Notation for the prototype assignment distributions and the precise form of the KL term should be introduced with an equation (or reference to a standard formulation) in the method section to improve reproducibility.
  2. [Experiments] The claim of 'minimal additional overhead (about 4% of fine-tuning time)' would be strengthened by reporting wall-clock times or FLOPs on the same hardware for both the pre-training stage and the subsequent fine-tuning stage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight opportunities to strengthen the abstract and clarify the experimental protocol. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] The abstract asserts that DASH 'consistently improves recognition under diverse noisy conditions' yet supplies no quantitative results, baselines, error bars, or experimental protocol. This absence is load-bearing for the central claim, as the soundness assessment notes the lack of evidence to substantiate the no-trade-off and generalization assertions.

    Authors: We agree that the abstract would be strengthened by including quantitative highlights. In the revised manuscript we will update the abstract to report key results from the LibriSpeech experiments (e.g., relative WER reductions on noisy test sets, preservation of clean accuracy, and a concise reference to the evaluation protocol and compute overhead). This directly addresses the concern about substantiating the central claims. revision: yes

  2. Referee: [Method / Experiments] The pre-training procedure relies on clean-noisy view pairs, but the manuscript does not state whether the noise types (or noise distribution) used during the label-free pre-training stage are disjoint from those appearing in the evaluation conditions. Overlap would allow the multi-layer distillation plus KL objective to encode corruption-specific statistics rather than invariant acoustic-to-semantic mappings, directly threatening the generalization claim to 'unseen' diverse noises.

    Authors: The referee correctly identifies that the manuscript does not explicitly state whether the noise distributions used for pre-training and evaluation are disjoint. We will revise the experimental section to provide a detailed description of the noise generation process, the specific noise types employed in each stage, and confirmation that the distributions are disjoint. This clarification will support the generalization claim by showing that the learned representations capture invariant mappings. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation chain absent from provided text

full rationale

The abstract and reader summary present DASH as a proposed self-distillation method with experimental results on LibriSpeech, but contain no equations, derivations, fitted parameters renamed as predictions, or self-citations that could reduce any claim to its own inputs by construction. The description of multi-layer distillation and KL minimization is stated at a high level without showing a load-bearing step that is self-definitional or imported via author overlap. This is the normal case of a self-contained empirical proposal whose central claim does not reduce to a fit or citation chain within the given material.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities detailed. The core premise relies on the unverified effectiveness of the proposed distillation for robustness.

axioms (1)
  • domain assumption Self-distillation from clean-noisy paired views via multi-layer representations and prototype KL divergence improves generalization to noise without labels or trade-offs.
    This is the central unproven premise enabling the label-free pre-training claim.

pith-pipeline@v0.9.1-grok · 5691 in / 1207 out tokens · 28380 ms · 2026-06-26T19:10:13.361819+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Introduction Recent advancements in automatic speech recognition (ASR) have been largely driven by the scaling up of model architec- tures and the utilization of massive training datasets, leading to significant performance improvements [1, 2, 3, 4]. Concur- rently, the evolution of speech foundation models has further propelled ASR capabilities by levera...

  2. [2]

    DASH: Dual-View Self-Distillation with Multi-Layer Hidden Representations for Robust Speech Recognition

    DASH 2.1. Clean-Noisy Pair with Dual Encoder Branches Figure 1 illustrates the overall architecture of DASH. The core intuition behind DASH is to encourage the model to learn noise- arXiv:2606.19203v1 [eess.AS] 17 Jun 2026 𝐷𝐷KL 𝑃𝑃clean Prototype score Subsampling stop gradient Encoder Layer 1 Encoder Layer 6 Encoder Layer 11 Encoder Layer 17 Projection Pr...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Dataset and Augmentation For training and dataset, we use LibriSpeech [18]train-960 for supervised ASR training and LibriLight Medium [21] for encoder-only self-distillation

    Experimental Setup 3.1. Dataset and Augmentation For training and dataset, we use LibriSpeech [18]train-960 for supervised ASR training and LibriLight Medium [21] for encoder-only self-distillation. For evaluation, we used greedy decoding without external language models and report WER on test-cleanandtest-otherof LibriSpeech. During self-distillation, th...

    2021

  4. [4]

    Experimental Results 4.1. Results on SNR Range Table 1 presents WER evaluations of the baseline, the stan- dard fine-tuning approaches, and the proposed DASH frame- work acrosstest-clean,test-other, and simulated noisy con- ditions. NOISEX-92 [27] was used for noise mixing, which consists of babble, pink, and white noise. Consistent with our preliminary h...

  5. [5]

    Conclusion In this paper, we proposed DASH, a dual-view self-distillation framework that improves ASR robustness by enforcing clean– noisy consistency. DASH adopts a decoupled two-stage pipeline: label-free encoder pre-training with an EMA teacher and prototype-based KL distillation across multiple encoder layers, followed by supervised fine-tuning. Exper...

  6. [6]

    Acknowledgments This work was partly supported by Institute of Informa- tion & communications Technology Planning & Evalua- tion(IITP) grant funded by the Korea government(MSIT)(RS- 2022-II220989, Development of Artificial Intelligence Tech- nology for Multi-speaker Dialog Modeling) and National Re- search Foundation of Korea(NRF) grant funded by the Ko- ...

    2022

  7. [7]

    Generative AI Use Disclosure Generative AI tools were used solely for editing and polishing the manuscript to improve linguistic clarity. These tools were not employed to produce any significant portion of the technical or scientific content, and the authors remain fully responsible and accountable for the integrity and final results of the work

  8. [8]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  9. [9]

    Fast Conformer With Linearly Scalable Atten- tion For Efficient Speech Recognition,

    D. Rekesh, N. R. Koluguri, S. Kriman, S. Majumdar, V . Noroozi, H. Huang, O. Hrinchuk, K. Puvvada, A. Kumar, J. Balam, and B. Ginsburg, “Fast Conformer With Linearly Scalable Atten- tion For Efficient Speech Recognition,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2023, pp. 1–8

    2023

  10. [10]

    V oxtral,

    A. H. Liu, A. Ehrenberg, A. Lo, C. Denoix, C. Barreau, G. Lam- ple, J.-M. Delignon, K. Raghavi Chandu, P. von Platen, P. R. Mud- direddy, and others, “V oxtral,”arXiv e-prints, pp. arXiv–2507, 2025

    2025

  11. [11]

    Qwen3-ASR Technical Re- port,

    X. Shi, X. Wang, Z. Guo, Y . Wang, P. Zhang, X. Zhang, Z. Guo, H. Hao, Y . Xi, B. Yang, and others, “Qwen3-ASR Technical Re- port,”arXiv e-prints, pp. arXiv–2601, 2026

    2026

  12. [12]

    Mockingjay: Unsupervised Speech Representation Learning with Deep Bidirectional Transformer Encoders,

    A. T. Liu, S.-w. Yang, P.-H. Chi, P.-c. Hsu, and H.-y. Lee, “Mockingjay: Unsupervised Speech Representation Learning with Deep Bidirectional Transformer Encoders,” inICASSP 2020- 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 6419–6423

    2020

  13. [13]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  14. [14]

    HuBERT: Self-Supervised Speech Rep- resentation Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “HuBERT: Self-Supervised Speech Rep- resentation Learning by Masked Prediction of Hidden Units,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, pp. 3451–3460, 2021

    2021

  15. [15]

    data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language,

    A. Baevski, W.-N. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, “data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language,” inProceedings of the 39th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvari, G. Niu, and S. Sabato, Eds., vol

  16. [16]

    2022, pp

    PMLR, Jul. 2022, pp. 1298–1312. [Online]. Available: https://proceedings.mlr.press/v162/baevski22a.html

    2022

  17. [17]

    WavLM: Large- Scale Self-Supervised Pre-Training for Full Stack Speech Pro- cessing,

    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 Pro- cessing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  18. [18]

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

    J.-B. Grill, F. Strub, F. Altch ´e, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Ghesh- laghi Azar, and others, “Bootstrap your own latent-a new ap- proach to self-supervised learning,”Advances in neural informa- tion processing systems, vol. 33, pp. 21 271–21 284, 2020

    2020

  19. [19]

    iBOT: Image BERT Pre-Training with Online Tok- enizer,

    J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong, “iBOT: Image BERT Pre-Training with Online Tok- enizer,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  20. [20]

    Emerging Properties in Self-Supervised Vision Transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bo- janowski, and A. Joulin, “Emerging Properties in Self-Supervised Vision Transformers,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), Oct. 2021, pp. 9650–9660

    2021

  21. [21]

    Speech Sim- CLR: Combining Contrastive and Reconstruction Objective for Self-supervised Speech Representation Learning,

    D. Jiang, W. Li, M. Cao, W. Zou, and X. Li, “Speech Sim- CLR: Combining Contrastive and Reconstruction Objective for Self-supervised Speech Representation Learning,” inProc. Inter- speech. ISCA, 2021, pp. 1544–1548

    2021

  22. [22]

    CR-CTC: Consistency Regulariza- tion 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 Regulariza- tion on CTC for Improved Speech Recognition,” inInternational Conference on Learning Representations (ICLR), 2025

    2025

  23. [23]

    HuBERT-VIC: Improving Noise- Robust Automatic Speech Recognition of Speech Foundation Model via Variance-Invariance-Covariance Regularization,

    H. Ahn, K. Jang, and H. Kim, “HuBERT-VIC: Improving Noise- Robust Automatic Speech Recognition of Speech Foundation Model via Variance-Invariance-Covariance Regularization,” inIn- terspeech 2025, 2025, pp. 3419–3423

    2025

  24. [24]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,

    A. Tarvainen and H. Valpola, “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., vol. 30. Curran Associates, Inc., 2017. [Online]. Ava...

    2017

  25. [25]

    Distilhubert: Speech Representation Learning by Layer-Wise Distillation of Hidden- Unit Bert,

    H.-J. Chang, S.-w. Yang, and H.-y. Lee, “Distilhubert: Speech Representation Learning by Layer-Wise Distillation of Hidden- Unit Bert,” inICASSP 2022 - 2022 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7087–7091

    2022

  26. [26]

    Lib- rispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An ASR corpus based on public domain audio books,” inICASSP 2015-2015 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2015, pp. 5206– 5210

    2015

  27. [27]

    Knowledge Distillation: A Good Teacher Is Pa- tient and Consistent,

    L. Beyer, X. Zhai, A. Royer, L. Markeeva, R. Anil, and A. Kolesnikov, “Knowledge Distillation: A Good Teacher Is Pa- tient and Consistent,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), Jun. 2022, pp. 10 925–10 934

    2022

  28. [28]

    Nature Machine In- telligence2(11), 665–673 (2020)

    R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge, and F. A. Wichmann, “Shortcut learning in deep neural networks,”Nature Machine Intelligence, vol. 2, no. 11, pp. 665–673, Nov. 2020. [Online]. Available: https: //doi.org/10.1038/s42256-020-00257-z

    work page doi:10.1038/s42256-020-00257-z 2020

  29. [29]

    Libri-Light: A Benchmark for ASR with Limited or No Su- pervision,

    J. Kahn, M. Rivi `ere, W. Zheng, E. Kharitonov, Q. Xu, P. Mazar´e, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegen, T. Likhoma- nenko, G. Synnaeve, A. Joulin, A. Mohamed, and E. Dupoux, “Libri-Light: A Benchmark for ASR with Limited or No Su- pervision,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA...

    2020

  30. [30]

    SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,” in Proc. Interspeech, ser. interspeech 2019. ISCA, Sep. 2019, pp. 2613–2617. [Online]. Available: http://dx.doi.org/10.21437/ Interspeech.2019-2680

    2019

  31. [31]

    MUSAN: A Music, Speech, and Noise Corpus

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A Music, Speech, and Noise Corpus,” 2015, arXiv:1510.08484. [Online]. Available: https://arxiv.org/abs/1510.08484

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  32. [32]

    Interspeech 2021 Deep Noise Suppression Challenge,

    C. K. A. Reddy, H. Dubey, K. Koishida, A. Nair, V . Gopal, R. Cutler, S. Braun, H. Gamper, R. Aichner, and S. Srinivasan, “Interspeech 2021 Deep Noise Suppression Challenge,” 2021, arXiv:2101.01902. [Online]. Available: https://arxiv.org/abs/ 2101.01902

    work page arXiv 2021

  33. [33]

    Efficient Sequence Transduction by Jointly Predicting Tokens and Durations,

    H. Xu, F. Jia, S. Majumdar, H. Huang, S. Watanabe, and B. Ginsburg, “Efficient Sequence Transduction by Jointly Predicting Tokens and Durations,” inProceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, Eds., vol

  34. [34]

    2023, pp

    PMLR, Jul. 2023, pp. 38 462–38 484. [Online]. Available: https://proceedings.mlr.press/v202/xu23g.html

    2023

  35. [35]

    InProceedings of the 23rd international conference on Machine learning (ICML ’06)

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” inProceedings of the 23rd International Conference on Machine Learning, ser. ICML ’06. New York, NY , USA: Association for Computing Machinery, 2006, pp. 369–376. [Online]. Available: https:/...

    work page doi:10.1145/1143844.1143891 2006

  36. [36]

    Assessment for automatic speech recognition: II. NOISEX-92: A database and an experiment to study the effect of additive noise on speech recognition systems,

    A. Varga and H. J. M. Steeneken, “Assessment for automatic speech recognition: II. NOISEX-92: A database and an experiment to study the effect of additive noise on speech recognition systems,”Speech Communication, vol. 12, no. 3, pp. 247–251, 1993. [Online]. Available: https://www.sciencedirect. com/science/article/pii/0167639393900953

    work page arXiv 1993