Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Noisy Disentanglement with Tri-stage Training for Noise-Robust Speech Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Inserting a 1.71M-parameter disentanglement module between encoder and decoder, trained with two losses under a tri-stage schedule, cuts noisy ASR word error by up to 25.7% on simulated and 10.6% on real CHiME-4 tests.

desk verdict Solid engineering paper with a plausible but not fully isolated claim; worth referee time, but the 360h results need a schedule control before they can be believed. read the letter →

arxiv 2509.01087 v1 pith:P4WGOSWR submitted 2025-09-01 eess.AS cs.SD

Shuangyuan Chen , Shuang Wei , Dongxing Xu , Yanhua Long This is my paper · ORCID
classification eess.AScs.SD
keywords noise-robustspeechrecognitionConformer-Transducernoisydisentanglementcleanrepresentationconsistencylossreconstructiontri-stagetrainingRNN-Tenhancement
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 sets out to show that end-to-end speech recognition can be made noise-robust by cleaning up internal representations rather than by adding a separate speech-enhancement front-end. It inserts a lightweight module between the Conformer encoder and the Transducer decoder that tries to split each noisy representation into a clean stream and a residual noise stream, sending only the clean stream on to the decoder. Two mean-squared-error losses guide the split: one pulls the clean stream toward the representation the same encoder produces for the corresponding clean utterance, and one requires the two streams together to reconstruct the original noisy representation. A three-stage schedule first trains the recognizer on clean speech, then trains the disentanglement module alone with everything else frozen, then fine-tunes the backbone. On LibriSpeech with CHiME-4 noise, the reported result is up to 25.7% relative word-error reduction on simulated noisy test sets, 10.6% on real CHiME-4 test sets, and no loss on clean test speech.

What carries the argument

The load-bearing mechanism is the NoisyD module: three four-layer MLPs (Encoder-C, Encoder-N, Decoder-CN) totaling 1.71M parameters, inserted between the Conformer encoder and the Transducer decoder. Encoder-C outputs the cleaned representation that feeds the decoder; Encoder-N extracts a residual noise representation; Decoder-CN reconstructs the original noisy representation from the concatenation. The reconstruction constraint makes the clean/noise split lossless, and the clean consistency loss L_CON anchors the clean branch to a frozen clean-trained encoder's representation via MSE. The tri-stage training schedule determines when each loss is active: clean pre-training establishes the rec

What would settle it

Replace the frozen clean encoder that produces the regression target with a differently initialized Conformer encoder of the same size, keep every other detail fixed, and retrain only stages two and three; if the CHiME-4 WER gains disappear, the benefit depends on the teacher representation space rather than on the disentanglement structure itself.

Watch

Extended reading notes

Core claim

The central claim is that a compact module between the Conformer encoder and the RNN-T decoder can make noisy internal representations close enough to clean representations that the decoder transcribes them with far fewer errors. Encoder-C extracts a cleaned representation for the decoder; Encoder-N extracts a noise-only stream; Decoder-CN reconstructs the original noisy representation from both. Loss L_CON pulls Encoder-C's output toward the frozen clean encoder's output for the same utterance; loss L_R forces the split to be information-complete. A tri-stage schedule pre-trains Conformer-T on clean speech, trains the module alone with the backbone frozen, then fine-tunes only the backbone.

Load-bearing premise

The load-bearing premise is that the frozen clean-trained Conformer encoder's representation space is one where mean-squared-error distance between noisy and clean representations tracks how much noise remains, so that minimizing L_CON suppresses noise without erasing linguistic information the decoder needs.

Editorial extensions

If this is right

  • The added 1.71M parameters are the only permanent cost at inference: the extra pre-trained encoder copy, Encoder-N, and Decoder-CN are discarded, so the method is a cheap retrofit to existing Conformer-Transducer systems.
  • Clean speech is not traded off: on the 360-hour setup, test-clean WER moved from 7.9% to 7.8%, and test-other from 16.7% to 14.3%, so the module can be inserted without a clean/noisy performance trade-off.
  • The gains transfer to noise types never seen in training: on DEMAND-based tests the model reported relative WER reductions of 12.4% at -5 dB and 25.5% at 15 dB.
  • The ablation results imply that both reconstruction (Encoder-N + Decoder-CN + L_R) and the tri-stage schedule are load-bearing; removing them leaves the system close to the baseline, and the schedule alone does not help.
  • Because the pre-trained clean encoder provides the regression target, the method inherits that encoder's representation quality as a ceiling on how clean the disentangled representation can become.

Reading between the lines

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

  • An implication the authors leave implicit: because the cleaning happens on encoder representations rather than waveforms, the same NoisyD recipe should transfer to other encoder-decoder ASR models, provided their encoders are stable enough to serve as regression targets.
  • The regression target comes from a frozen clean-trained encoder, so the method's ceiling is set by that teacher's representation space; upgrading the encoder later would require regenerating the paired targets and re-running stages two and three.
  • A cheap practical extension would generate the clean-noisy pairs on the fly with any noise-corruption pipeline, eliminating the need for a fixed paired corpus and potentially widening the SNR range covered during training.
  • It is an open, testable question whether representation-level MSE alignment is more ASR-friendly than waveform-level speech enhancement; a direct comparison at equal parameter count would clarify where the benefit comes from.
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 NoisyD-CT, a Conformer-Transducer ASR system augmented with a lightweight NoisyD module inserted between the Conformer encoder and the RNN-T decoder. NoisyD consists of Encoder-C, Encoder-N, and Decoder-CN. Two auxiliary losses are added: a clean-representation consistency loss L_CON that aligns Encoder-C's output with representations from a frozen clean-only pretrained encoder, and a noisy reconstruction loss L_R that requires the concatenation of Encoder-C and Encoder-N outputs to reconstruct the noisy encoder representation. Training uses a tri-stage schedule: clean pretraining, noisy-disentanglement training with the backbone frozen, and fine-tuning of the backbone. Experiments on LibriSpeech (100h and 360h) with simulated CHiME-4 noise, real CHiME-4 test sets, and unseen DEMAND noise show consistent WER gains over Conformer-Transducer baselines, with headline relative reductions of 25.7% on simulated noisy speech and 10.6% on CHiME-4, plus small clean-test improvements.

Significance. The contribution is practical and the experimental design is above the minimum bar: the paper includes a schedule-only control (B3+Tri) at 100h, ablations that separate L_CON from the full reconstruction path (S4 vs S3), matched/unseen noise conditions, and a modest parameter overhead (1.71M). If the results are reproducible, NoisyD-CT is a useful add-on for noise-robust E2E ASR without external enhancement front-ends. The main weaknesses are statistical (single runs, no intervals) and the absence of a schedule-only control at the 360h scale where the strongest claims are made. The 'disentanglement' interpretation also goes beyond what the losses formally enforce. These are fixable within the paper's scope, so I do not view them as fatal.

major comments (3)
  1. [Section 6.4, Table 5] The 360-hour experiments compare S3+ (NoisyD + tri-stage) with B3+ (Conformer-T trained from scratch on combined clean+noisy data). These systems differ in both architecture and training schedule. The only control that isolates the schedule, B3+Tri, is reported at 100h (Table 2) and actually performs slightly worse than B3. Because schedule effects can interact with dataset size and noise diversity, the headline 360h gains (e.g., 25.7% relative at SNR=15) are not fully attributable to the NoisyD module. Please run B3+Tri on the 360h condition (or otherwise control for schedule at that scale) and report it alongside Table 5.
  2. [Section 6, Tables 3-7] All WER results are single runs with no confidence intervals or significance tests. Several key differences are small (test-clean 7.9 to 7.8 in Table 5; CHiME-4 BUS 39.6 to 38.3 in Table 6) and could lie within run-to-run variability. Please provide repeated-seed means and standard deviations, or a significance test, for the main B3+ vs S3+ comparisons and the CHiME-4 results.
  3. [Section 4.2, Eqs. (7)-(8)] The claims that Encoder-N extracts 'pure noise' and that the module performs 'disentanglement' are stronger than the losses enforce. L_R only requires that the concatenation [tilde h_clean; tilde h_noisy] reconstructs ht-noisy; it does not by itself prevent Encoder-N from encoding residual speech information or prevent redundancy between the two branches. Please add an explicit disentangling constraint (e.g., orthogonality or information-reduction) or soften the interpretation, and ideally add an analysis showing that Encoder-N captures noise rather than content (e.g., an ablation that replaces or removes Encoder-N output).
minor comments (5)
  1. [Eqs. (7)-(8)] The formulas are written as if L_CON and L_R were simple differences; the text says MSE. Please write the squared L2 norm explicitly.
  2. [Table 3] The checkmark notation is difficult to read, especially whether L_CON+LR are active during fine-tuning when the row shows ND frozen. Please clarify with a separate column or footnote.
  3. [Section 6.5, Figure 4] The visualization is partly circular because L_CON explicitly trains tilde h_clean to match ht. It is a useful sanity check but not independent evidence of disentanglement.
  4. [Throughout] Typos and grammar: 'a especially designed', 'mechanisim', 'with the model size remains unchanged', 'the whole model loss function'. Also, Eqs. (7)-(8) use inconsistent notation for the reconstruction target.
  5. [Sections 2 and 6] The related work lists SE-based and SSL-based robust ASR methods, but no experimental comparison with such systems is included. A short comparison (or an explicit statement that this is out of scope) would help position the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central WER claims are evaluated on held-out test sets, and the proposed losses are training objectives rather than fitted predictions.

full rationale

The paper's central claim is that NoisyD-CT, trained with L_CON (Eq. 7) and L_R (Eq. 8) under a tri-stage schedule, improves ASR WER on noisy speech. These WER numbers are measured on held-out simulated LibriSpeech-noisy test sets, held-out CHiME-4 real-noise sets, and unseen DEMAND noise, so the headline result is not defined into existence by the training loss. L_CON is an MSE regression target produced by a frozen clean-pretrained encoder; it is a training signal, not a test-set fit. L_R is likewise a reconstruction auxiliary loss. Neither loss is renamed as a 'prediction' of the reported WER improvements. The paper contains no load-bearing self-citations: the cited [38], [39], [40], [41] are standard external references for Conformer, Transducer, and CTC/RNN-T losses, and no uniqueness theorem is imported from the authors' prior work. The proposed architecture is not justified by citing an ansatz; the module and losses are specified in Section 4.2 and ablated in Table 3. The Figure 4 visualization does show that the output of Encoder-C matches the clean target ht, but this closeness is directly enforced by L_CON, so it is a by-construction check of the training objective rather than independent evidence of disentanglement; however, this observation does not undermine the primary WER-based evaluation. The main experimental limitation is the absence of a B3+Tri control at the 360-hour scale, which would better isolate the contribution of the tri-stage schedule from the NoisyD module itself; this is an attribution/control concern, not a circularity, because the reported gains are still measured on held-out data. Overall, no derivation step reduces to its own inputs by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central contribution rests on fitted loss weights, a hand-chosen SNR range, and strong assumptions about the geometry of the Conformer representation space. The clean and noise representations are internal constructs that are only weakly validated. The external WER benchmark provides independent evidence, but the module's internal decomposition is not directly proven.

free parameters (6)
  • loss weight alpha = 0.3
    Weight for L_CON in Eq. 9; set to 0.3 without sensitivity analysis, affects the balance between representation alignment and ASR loss.
  • loss weight beta = 1
    Weight for L_R in Eq. 9; set to 1 with no ablation reported.
  • CTC loss weight mu = 0.3
    Weight for L_CTC in Eq. 5; standard value, not tuned in this paper.
  • RNN-T loss weight gamma = 1
    Weight for L_RNN-T in Eq. 5; standard value, not tuned.
  • SNR range for data simulation = -5 to 15 dB
    Random SNRs in this range were used to synthesize noisy training data; the range is chosen by hand.
  • NoisyD module architecture = 1.71M parameters
    Four linear layers with ReLU/Sigmoid were chosen to keep parameter overhead small; no architecture search or ablation on depth/width.
assumptions (4)
  • domain assumption The frozen clean Conformer encoder's representation space is such that MSE alignment between noisy and clean representations improves downstream ASR utility.
    The L_CON loss in Eq. 7 assumes feature-space closeness is a valid proxy for noise suppression that helps the decoder. No analysis of the representation geometry is provided.
  • domain assumption Noise can be separated from speech in the top Conformer representation using a small MLP encoder-decoder.
    The NoisyD module architecture (four linear layers) presupposes a simple, linearly-transformable separation of noise and speech in the high-level feature space.
  • domain assumption Synthetic additive noise mixtures at random SNRs capture the distribution of real-world noise conditions.
    All paired clean-noisy training data are generated by mixing CHiME-4 noise with LibriSpeech speech. Generalization to real CHiME-4 is tested, but the simulation is still a strong assumption about acoustic diversity.
  • ad hoc to paper The reconstruction loss L_R forces Encoder-N to extract only noise, not speech content.
    There is no direct verification that the noise representation is free of linguistic content; the claim is inferred from the reconstruction objective and downstream WER gains.
invented entities (2)
  • Noise representation \tilde{h}_noisy from Encoder-N
    purpose: Purported pure-noise component used as input to Decoder-CN for noisy reconstruction.
    No visualization or metric shows this representation contains only noise; its behavior is inferred from the reconstruction loss. It is a latent variable with no external validation.
  • Disentangled clean representation \tilde{h}_clean independent evidence
    purpose: Purified representation fed to the Transducer Decoder for ASR.
    Visualized in Figure 4 for a single utterance and indirectly supported by WER improvements across test sets. The existence of a clean representation is plausible but the disentanglement claim is stronger than the evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Noisy Disentanglement with Tri-stage Training for Noise-Robust Speech Recognition." pith.science (2026). https://pith.science/paper/P4WGOSWR

@misc{pith2026250901087,
  author       = {Pith},
  title        = {Pith review of: Noisy Disentanglement with Tri-stage Training for Noise-Robust Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P4WGOSWR}},
  note         = {Machine review of arXiv:2509.01087}
}
read the original abstract

To enhance the performance of end-to-end (E2E) speech recognition systems in noisy or low signal-to-noise ratio (SNR) conditions, this paper introduces NoisyD-CT, a novel tri-stage training framework built on the Conformer-Transducer architecture. The core of NoisyD-CT is a especially designed compact noisy disentanglement (NoisyD) module (adding only 1.71M parameters), integrated between the Conformer blocks and Transducer Decoder to perform deep noise suppression and improve ASR robustness in challenging acoustic noise environments. To fully exploit the noise suppression capability of the NoisyD-CT, we further propose a clean representation consistency loss to align high-level representations derived from noisy speech with those obtained from corresponding clean speech. Together with a noisy reconstruction loss, this consistency alignment enables the NoisyD module to effectively suppress noise while preserving essential acoustic and linguistic features consistent across both clean and noisy conditions, thereby producing cleaner internal representations that enhance ASR performance. Moreover, our tri-stage training strategy is designed to fully leverage the functionalities of both the noisy disentanglement and speech recognition modules throughout the model training process, ultimately maximizing performance gains under noisy conditions. Our experiments are performed on the LibriSpeech and CHiME-4 datasets, extensive results demonstrate that our proposed NoisyD-CT significantly outperforms the competitive Conformer-Transducer baseline, achieving up to 25.7% and 10.6% relative word error rate reductions on simulated and real-world noisy test sets, respectively, while maintaining or even improving performance on clean speech test sets. The source code, model checkpoint and data simulation scripts will be available at https://github.com/litchimo/NoisyD-CT.

Figures

Figures reproduced from arXiv: 2509.01087 by the authors.

Figure 1
Figure 1. The framework of Conformer-Transducer. ity with self-supervised pretraining’s noise robustness, miti￾gating information loss and speech distortion. However, this method is limited to offline ASR tasks and is unsuitable for real-time applications. To further refine integration, [35] in￾troduce a knowledge distillation-based joint training method, which constructs a pipeline where the ASR and SE models are treated as … view at source ↗
Figure 2
Figure 2. The framework of proposed NoisyD-CT E2E ASR. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Block diagram of tri-stage training strategy on the NoisyD-CT framework. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of the (a) noisy representation [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 45 canonical work pages

  1. [1]

    J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, Y . Bengio, Attention- based models for speech recognition, in: Proc. NeurIPS, 2015, pp. 577– 585. 10

    work page 2015

  2. [2]

    B. Li, A. Gulati, J. Yu, A better and faster end-to-end model for streaming ASR, in: Proc. ICASSP, 2021, pp. 5634–5638

    work page 2021

  3. [3]

    S. Kim, T. Hori, S. Watanabe, Joint CTC-attention based end-to-end speech recognition using multi-task learning, in: Proc. ICASSP, 2017, pp. 4835–4839

    work page 2017

  4. [4]

    L. Dong, S. Xu, B. Xu, Speech-transformer: A no-recurrence sequence- to-sequence model for speech recognition, in: Proc. ICASSP, 2018, pp. 5884–5888

    work page 2018

  5. [5]

    Hinton, L

    G. Hinton, L. Deng, D. Yu, Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups, IEEE Signal processing magazine, V ol. 29, 2012, pp. 82–97

    work page 2012

  6. [6]

    J. H. Martin, D. Jurafsky, Speech and language processing: An introduc- tion to natural language processing, computational linguistics, and speech recognition, V ol. 23, Pearson/Prentice Hall Upper Saddle River, 2009

    work page 2009

  7. [7]

    Xiong, L

    W. Xiong, L. Wu, F. Alleva, J. Droppo, X. Huang, A. Stolcke, The mi- crosoft 2017 conversational speech recognition system, in: Proc. ICASSP, 2018, pp. 5934–5938

    work page 2017

  8. [8]

    Li, et al., Recent advances in end-to-end automatic speech recogni- tion, APSIPA Transactions on Signal and Information Processing, V ol

    J. Li, et al., Recent advances in end-to-end automatic speech recogni- tion, APSIPA Transactions on Signal and Information Processing, V ol. 11, 2022

    work page 2022

Show all 46 references
  1. [9]

    M. L. Seltzer, D. Yu, Y . Wang, An investigation of deep neural networks for noise robust speech recognition, in: Proc. ICASSP, 2013, pp. 7398– 7402

    2013

  2. [10]

    Weninger, H

    F. Weninger, H. Erdogan, S. Watanabe, Speech enhancement with LSTM recurrent neural networks and its application to noise-robust ASR, in: Proc. LV A/ICA, 2015, pp. 91–99

    2015

  3. [11]

    Scalart, et al., Speech enhancement based on a priori signal to noise estimation, in: Proc

    P. Scalart, et al., Speech enhancement based on a priori signal to noise estimation, in: Proc. ICASSP, 1996, pp. 629–632

    1996

  4. [12]

    Pandey, D

    A. Pandey, D. Wang, A new framework for CNN-based speech enhance- ment in the time domain, IEEE/ACM Transactions on Audio, Speech, and Language Processing, V ol. 27, 2019, pp. 1179–1188

    2019

  5. [13]

    Defossez, G

    A. Defossez, G. Synnaeve, Y . Adi, Real time speech enhancement in the waveform domain, in: Proc. Interspeech, 2020, pp. 3291–3295

    2020

  6. [14]

    Michelsanti, Z

    D. Michelsanti, Z. Tan, Conditional generative adversarial networks for speech enhancement and noise-robust speaker verification, in: Proc. In- terspeech, 2017, pp. 2008–2012

    2017

  7. [15]

    Z. Wang, P. Wang, D. Wang, Complex spectral mapping for single-and multi-channel speech enhancement and robust ASR, IEEE /ACM trans- actions on audio, speech, and language processing, V ol. 28, 2020, pp. 1778–1787

    2020

  8. [16]

    Fujimoto, H

    M. Fujimoto, H. Kawai, One-pass single-channel noisy speech recogni- tion using a combination of noisy and enhanced features, in: Proc. Inter- speech, 2019, pp. 486–490

    2019

  9. [17]

    Z. Wang, D. Wang, A joint training framework for robust automatic speech recognition, IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, V ol. 24, 2016, pp. 796–806

    2016

  10. [18]

    C. Fan, J. Yi, J. Tao, Gated recurrent fusion with joint training frame- work for robust end-to-end speech recognition, IEEE/ACM Transactions on Audio, Speech, and Language Processing, V ol. 29, 2020, pp. 198–209

    2020

  11. [19]

    Donahue, B

    C. Donahue, B. Li, R. Prabhavalkar, Exploring speech enhancement with generative adversarial networks for robust speech recognition, in: Proc. ICASSP, 2018, pp. 5024–5028

    2018

  12. [20]

    Pandey, C

    A. Pandey, C. Liu, Y . Wang, Y . Saraf, Dual application of speech enhance- ment for automatic speech recognition, in: Proc. SLT, 2021, pp. 223–228

    2021

  13. [21]

    Kinoshita, T

    K. Kinoshita, T. Ochiai, M. Delcroix, T. Nakatani, Improving noise robust automatic speech recognition with single-channel time-domain enhance- ment network, in: Proc. ICASSP, 2020, pp. 7009–7013

    2020

  14. [22]

    Dubey, A

    H. Dubey, A. Aazami, V . Gopal, et al., ICASSP 2023 deep noise suppres- sion challenge, IEEE Open Journal of Signal Processing, 2024

    2023

  15. [23]

    Y . Ju, J. Chen, S. Zhang, TEA-PSE 3.0: Tencent-Ethereal-Audio-Lab per- sonalized speech enhancement system for ICASSP 2023 DNS-Challenge, in: Proc. ICASSP, 2023, pp. 1–2

    2023

  16. [24]

    B. Liu, S. Nie, S. Liang, W. Liu, M. Yu, L. Chen, S. Peng, C. Li, Jointly adversarial enhancement training for robust end-to-end speech recogni- tion, in: Proc. Interspeech, 2019, pp. 491–495

    2019

  17. [25]

    D. Ma, N. Hou, H. Xu, E. S. Chng, et al., Multitask-based joint learn- ing approach to robust ASR for radio communication speech, in: Proc. APSIPA ASC, 2021, pp. 497–502

    2021

  18. [26]

    Z. Li, C. Liang, T. Lohrenz, M. Sach, B. M ¨oller, T. Fingscheidt, An effi- cient and noise-robust audiovisual encoder for audiovisual speech recog- nition, in: Proc. Interspeech, 2023, pp. 1583–1587

    2023

  19. [27]

    Y . Wang, J. Li, H. Wang, Y . Qian, C. Wang, Y . Wu, Wav2vec-switch: Contrastive learning from original-noisy speech pairs for robust speech recognition, in: Proc. ICASSP, 2022, pp. 7097–7101

    2022

  20. [28]

    Baevski, Y

    A. Baevski, Y . Zhou, A. Mohamed, M. Auli, wav2vec 2.0: A framework for self-supervised learning of speech representations, in: Proc. NeurIPS, V ol. 33, 2020, pp. 12449–12460

    2020

  21. [29]

    Q. Zhu, J. Zhang, Z. Zhang, M. Wu, X. Fang, L. Dai, A noise-robust self- supervised pre-training model based speech representation learning for automatic speech recognition, in: Proc. ICASSP, 2022, pp. 3174–3178

    2022

  22. [30]

    Huang, Z

    W. Huang, Z. Zhang, Y . T. Yeung, X. Jiang, Q. Liu, SPIRAL: Self- supervised perturbation-invariant representation learning for speech pre- training, in: Proc. ICLR, 2022

    2022

  23. [31]

    H. Wang, Y . Qian, X. Wang, Improving noise robustness of contrastive speech representation learning with speech reconstruction, in: Proc. ICASSP, 2022, pp. 6062–6066

    2022

  24. [32]

    D. Ng, R. Zhang, J. Q. Yip, Z. Yang, De’hubert: Disentangling noise in a self-supervised model for robust speech recognition, in: Proc. ICASSP, 2023, pp. 1–5

    2023

  25. [33]

    Ravanelli, J

    M. Ravanelli, J. Zhong, S. Pascual, Multi-task self-supervised learning for robust speech recognition, in: Proc. ICASSP, 2020, pp. 6989–6993

    2020

  26. [34]

    Q. Zhu, J. Zhang, Z. Zhang, L. Dai, A joint speech enhancement and self- supervised representation learning framework for noise-robust speech recognition, in: IEEE /ACM Transactions on Audio, Speech, and Lan- guage Processing, V ol. 31, 2023, pp. 1927–1939

    2023

  27. [35]

    G. W. Lee, H. K. Kim, D. Kong, Knowledge distillation-based training of speech enhancement for noise-robust automatic speech recognition, IEEE Access, V ol. 12, 2024, pp. 72707–72720

    2024

  28. [36]

    H. Hu, T. Tan, Y . Qian, Generative adversarial networks based data aug- mentation for noise robust speech recognition, in: Proc. ICASSP, 2018, pp. 5044–5048

    2018

  29. [37]

    Goodfellow, J

    I. Goodfellow, J. PougetAbadie, M. Mirza, Generative adversarial nets, Proc. NeurIPS, V ol. 27, 2014

    2014

  30. [38]

    Gulati, J

    A. Gulati, J. Qin, C. Chiu, N. Parmar, Conformer: convolution- augmented transformer for speech recognition, in: Proc. Interspeech, 2020, pp. 5036–5040

    2020

  31. [39]

    Zhang, H

    Q. Zhang, H. Lu, H. Sak, Transformer Transducer: A streamable speech recognition model with Transformer encoders and RNN-T Loss, in: Proc. ICASSP, 2020, pp. 7829–7833

    2020

  32. [40]

    Graves, Sequence transduction with recurrent neural networks, in: Proc

    A. Graves, Sequence transduction with recurrent neural networks, in: Proc. ICML, 2012

    2012

  33. [41]

    J. Jeon, E. Kim, Multitask learning and joint optimization for transformer- RNN-transducer speech recognition, in: Proc. ICASSP, 2021, pp. 6793– 6797

    2021

  34. [42]

    Panayotov, G

    V . Panayotov, G. Chen, D. Povey, S. Khudanpur, Librispeech: an ASR corpus based on public domain audio books, in: Proc. ICASSP, 2015, pp. 5206–5210

    2015

  35. [43]

    Vincent, S

    E. Vincent, S. Watanabe, A. A. Nugraha, J. Barker, R. Marxer, An anal- ysis of environment, microphone and data simulation mismatches in ro- bust speech recognition, Computer Speech & Language, V ol. 46, 2017, pp. 535–557

    2017

  36. [44]

    Thiemann, N

    J. Thiemann, N. Ito, E. Vincent, DEMAND: A Collection of Multi- channel Recordings of Acoustic Noise in Diverse Environments, Zenodo (2013)

    2013

  37. [45]

    D. S. Park, W. Chan, Y . Zhang, C. Chiu, B. Zoph, E. D. Cubuk, Q. V . Le, Specaugment: A simple data augmentation method for automatic speech recognition, in: Proc. Interspeech, 2019, pp. 2613–2617

    2019

  38. [46]

    Sennrich, B

    R. Sennrich, B. Haddow, A. Birch, Neural machine translation of rare words with subword units, arXiv preprint arXiv:1508.07909, 2015. 11

    2015 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.