Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Where Speech Enhancement Hurts Recognition: An Inference Time Polar Projection Diagnosis

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Speech enhancement hurts ASR mainly through magnitude strength, not phase correction, and the right strength depends on the recognizer.

desk verdict Clean, usable diagnosis: magnitude strength is the operative SE–ASR axis, estimated phase does not help, and optimal alpha is recognizer-dependent; single-corpus STFT scope is the real limit, not a hidden flaw. read the letter →

arxiv 2607.11157 v1 pith:ZUYBKRDE submitted 2026-07-13 eess.AS

Mingyue Huo , Yuheng Zhang , Hao Zhang This is my paper · ORCID
classification eess.AS
keywords speechenhancementrecognitionmagnitudeandphasepolarprojectionSTFTmaskover-suppressionwav2vec2.0Whisper
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

Speech enhancement can make audio sound cleaner to people while still raising word error for automatic speech recognition. The usual explanations—artifacts and over-suppression—are qualitative and do not say which part of the enhancer is to blame. This paper introduces inference-time polar projection: any STFT-domain complex mask is rewritten so its magnitude strength and phase correction can be dialed independently without retraining. Sweeping those two knobs on frozen enhancers and recognizers shows that magnitude strength is the operative axis, while the enhancer’s estimated phase correction gives no recognition benefit. The best magnitude setting is recognizer-dependent: waveform-input wav2vec 2.0 wants strong correction, while log-Mel, noise-robust Whisper prefers a milder one. The same knob also serves as a training-free repair that can be calibrated on a small validation set for voice assistants that share enhanced speech between human listeners and machines.

What carries the argument

Inference-time polar projection: given a complex mask M = A e^{jϕ}, form M_{α,γ} = A^α e^{jγϕ} and apply it to the noisy STFT. α scales magnitude correction toward identity; γ scales phase rotation toward zero. Sweeping the pair on frozen models turns SE–ASR mismatch into measurable magnitude and phase effects and also supplies a training-free magnitude calibration.

What would settle it

On a different noise or reverberation corpus, or with a non-STFT enhancer whose input–output map cannot be treated as a complex ratio, show that sweeping phase correction γ systematically lowers word error while magnitude strength α does not, or that both recognizers share the same optimal α.

Watch

Extended reading notes

Core claim

When an STFT-domain enhancer’s complex mask is factored into independent magnitude strength α and phase-correction strength γ, ASR degradation is driven by magnitude strength, not by the estimated phase correction. Reusing the noisy phase is as good as or better than applying the enhancer’s phase, while the optimal α is recognizer-dependent: wav2vec 2.0 improves nearly monotonically with stronger magnitude correction, whereas Whisper is U-shaped and prefers weaker correction.

Load-bearing premise

The claim rests on the idea that sweeping polar-projected masks (and effective masks recovered after the fact) on one standard noisy-speech corpus with two frozen public enhancers and two public ASR models is enough to identify the operative axis of SE-induced recognition harm more broadly.

Editorial extensions

If this is right

  • Any STFT-mask enhancer can be turned into an ASR-aware front end by calibrating a single magnitude strength α on a small validation set, with no retraining of enhancer or recognizer.
  • Systems that serve both human listeners and machines can switch α between a perceptual operating point and a recognition operating point at inference time.
  • Phase-aware complex-mask training that improves listening metrics need not be assumed to help recognition; magnitude-oriented training or post-hoc magnitude scaling is the more direct lever.
  • Per-utterance α selection is an upper-bound opportunity: oracle utterance-level choice cuts WER far more than a single global α, motivating lightweight predictors of magnitude strength.

Reading between the lines

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

  • If magnitude is the main mismatch axis, joint SE–ASR training may be overkill for many deployments; a frozen enhancer plus a cheap α controller could capture most of the gain.
  • Recognizers that already train on noisy log-Mel features may systematically prefer weaker front-end suppression than waveform-input models, suggesting that front-end strength should be conditioned on the recognizer’s input representation.
  • The same polar-projection diagnostic could be used to audit whether generative or time-domain enhancers that only admit an effective mask still over-correct magnitude relative to a target ASR.
  • Voice-assistant pipelines that currently interpolate enhanced and noisy waveforms may be able to replace that single coupled knob with an explicit magnitude-only α control and obtain clearer diagnostics.
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 / 6 minor

Summary. The paper studies why speech enhancement that improves perceptual quality can still raise ASR word error rate. It introduces inference-time polar projection of an STFT-domain complex (or effective) mask M=A e^{jφ} into M_{α,γ}=A^α e^{jγφ}, with α controlling magnitude strength and γ controlling phase correction. Sweeping these parameters on frozen SE and ASR models is used to localize degradation. On VoiceBank+DEMAND with FRCRN-SE and MossFormerGAN-SE, and with Whisper-large-v3 and wav2vec 2.0, the authors report that magnitude strength is the operative axis, estimated phase correction gives no recognition benefit (γ=0 is best or near-best), and the optimal α is recognizer-dependent: wav2vec 2.0 prefers strong correction while Whisper prefers milder correction. Polar projection is also proposed as a training-free mitigation by calibrating α for a target recognizer.

Significance. If the reported dissociation holds, the work converts qualitative SE–ASR explanations (artifacts, over-suppression) into a simple, measurable two-axis diagnosis at the mask level, complementary to output-level error decompositions and observation-adding. Strengths include controlled α/γ sweeps with frozen models, paired bootstrap significance (B=10^4), SNR-wise breakdowns (Table III), and generalization from a strict cIRM model to an effective-mask analysis of MossFormerGAN-SE (Fig. 4). The recognizer dependence (waveform vs log-Mel / noise-robust training) is a concrete, falsifiable finding with practical deployment implications for voice assistants. The method is training-free and directly applicable to STFT-mask front ends. Scope is appropriately limited to STFT-domain enhancement and the evaluated estimated phases, which the Limitations section states clearly.

major comments (3)
  1. The diagnostic claims (Figs. 2–4, Table III) are supported by controlled sweeps, but the mitigation claim—that α can be calibrated on a small validation set for a target recognizer without retraining (Abstract, Contributions, §VII)—is not demonstrated with a proper held-out protocol. Optimal α* values appear selected on the same evaluation set used to report WER. A short experiment (e.g., tune α on a validation split, report WER on a disjoint test split, and compare to noisy / full-mask / observation-adding baselines) is needed to substantiate the mitigation contribution as stated.
  2. §VI-D recovers an effective mask as M̃=Ŝ⊘X with flooring of |X| and clipping of |M̃|. The paper should quantify sensitivity of the α/γ conclusions to the floor ε and clip range (or report that conclusions are stable under a stated range). Without this, the claim that the diagnosis is not tied to a strict complex-mask form rests on a post-hoc inversion whose numerical choices are under-specified.
  3. Table II establishes prevalence of SE–ASR mismatch for six enhancers with Whisper, but polar projection is applied only to FRCRN-SE and MossFormerGAN-SE. That is reasonable for non-mask generative systems, yet the Abstract and §I frame the diagnosis as applying to “any SE front end in the STFT mask domain.” Explicitly restricting the operative-axis claim to the two analyzed systems (or adding one more mask-based enhancer) would better match the evidence to the wording.
minor comments (6)
  1. Throughout (e.g., §IV-A, Table II): “V oiceBank” appears with a spurious space; fix typesetting of VoiceBank+DEMAND and similar names (caf´e, etc.).
  2. Eq. (1) and footnote 1: α,γ∈[0,1] is clear, but briefly state whether α is applied in linear magnitude or whether any per-bin clipping of A^α is used when A>1 or A≪1.
  3. Fig. 2–4: mark noisy (α=0) and full-mask (α=1) WER as horizontal reference lines and report 95% bootstrap CIs on the curves or at α* to match the significance language in the text.
  4. Table I is helpful; add a one-line pointer in §III that OA corresponds to the linear path M_λ=λM+(1−λ)1 so readers see the contrast with polar projection immediately next to Eq. (1).
  5. §VII Limitations correctly scopes the phase result to estimated corrections of the evaluated enhancers; consider repeating that caveat once in the Abstract/Conclusion so the “phase provides no recognition benefit” phrasing is not over-read as a claim about phase information in general.
  6. Minor consistency: abstract uses “wav2vec2.0” vs body “wav2vec 2.0”; unify naming and checkpoint identifiers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: polar projection is an empirical intervention with free sweep parameters; WER optima are measured, not forced by construction.

full rationale

The paper's central claims (magnitude strength is the operative axis of SE-ASR degradation; estimated phase correction provides no recognition benefit; optimal magnitude strength is recognizer-dependent) are obtained by applying free parameters α and γ to frozen SE masks, reconstructing waveforms, and measuring WER on VoiceBank+DEMAND with frozen ASR models. Equation (1) defines M_{α,γ}=A^α e^{jγφ} as a controllable intervention, not a fit to a target WER; α* is selected by the measured minimum WER (Figs. 2–4, Table III), with paired bootstrap significance. Relation to observation adding is shown only to contrast a coupled path with the two-axis diagnosis; it does not redefine success. Effective-mask recovery for MossFormerGAN-SE is a post-hoc ratio Œ=Ŝ⊘X used for the same sweep, not a circular redefinition of the enhancer. No self-definitional loop, fitted-input-called-prediction, load-bearing self-citation uniqueness claim, or renaming of a known result appears. The work is self-contained empirical diagnosis against external public models and a standard corpus; score 0 is appropriate.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The paper is empirical. Load-bearing background is standard STFT complex-mask algebra and the assumption that intervening on magnitude and phase via A^alpha e^{j gamma phi} isolates the effects of interest. Free parameters are the sweep coordinates and the chosen operating points alpha-star; no new physical entities are postulated.

free parameters (2)
  • alpha (magnitude strength) = recognizer-dependent (e.g. ~0.25 Whisper, ~0.85 wav2vec 2.0 on FRCRN)
    Continuous control in [0,1] swept at inference; optimal alpha-star is selected by measured WER on the evaluation set (or per-SNR/per-utterance oracles).
  • gamma (phase-correction strength) = 0 (optimal in nearly all settings)
    Continuous control in [0,1] swept at inference; consistently found near 0 for best WER.
assumptions (3)
  • domain assumption A complex STFT mask M = A e^{j phi} applied as S-hat = M X fully captures the enhancer’s magnitude and phase action for diagnostic purposes.
    Invoked in Sec. III and Table I; standard for cIRM enhancers and extended post-hoc to effective masks.
  • ad hoc to paper Restricting alpha, gamma to [0,1] (between identity and full mask) is sufficient to diagnose over-suppression and phase effects.
    Stated in Sec. III footnote; alpha > 1 is not explored.
  • domain assumption WER differences under paired bootstrap over utterances are a valid measure of recognition impact of mask interventions.
    Used throughout Sec. VI for significance claims.
invented entities (1)
  • inference-time polar projection M_{alpha,gamma} = A^alpha e^{j gamma phi} independent evidence
    purpose: Factorize any STFT complex (or effective) mask into independent magnitude-strength and phase-correction controls for diagnosis and mitigation.
    Defined in Eq. (1) and Fig. 1; it is a methodological construct, not a physical entity. Independent evidence is the empirical WER surfaces it produces.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Where Speech Enhancement Hurts Recognition: An Inference Time Polar Projection Diagnosis." pith.science (2026). https://pith.science/paper/ZUYBKRDE

@misc{pith2026260711157,
  author       = {Pith},
  title        = {Pith review of: Where Speech Enhancement Hurts Recognition: An Inference Time Polar Projection Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUYBKRDE}},
  note         = {Machine review of arXiv:2607.11157}
}
abstract

Speech enhancement (SE) can substantially improve perceptual quality, yet enhanced speech does not necessarily improve automatic speech recognition (ASR). Existing remedies, such as retraining the enhancer jointly with recognizer or interpolating enhanced speech with the noisy input, can mitigate this mismatch, but common explanations such as artifacts and over-suppression remain qualitative and do not localize which enhancement component harms recognition. We propose inference time polar projection, a diagnosis for STFT domain enhancement. Given a mask $M=Ae^{j\phi}$, polar projection forms $M_{\alpha,\gamma}=A^\alpha e^{j\gamma\phi}$, where $\alpha$ controls magnitude strength and $\gamma$ controls phase correction. Sweeping these controls on frozen SE and ASR models turns ASR degradation into measurable magnitude and phase effects. Our projection analysis shows that magnitude strength is the operative axis, while estimated phase correction provides no recognition benefit. The optimal magnitude strength is recognizer dependent: waveform-input wav2vec2.0 favors strong correction, whereas log-Mel-input, noise-robust Whisper prefers weaker correction. Finally, the projection provides a simple mitigation for any SE front end in the STFT mask domain, without retraining either the enhancer or the recognizer, making it directly useful for voice assistants and agents that rely on enhanced speech.

Figures

Figures reproduced from arXiv: 2607.11157 by the authors.

Figure 1
Figure 1. Inference-time polar projection in the complex mask [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Magnitude sweep (γ=0): WER vs. magnitude scaling strength α. While wav2vec 2.0 improves monotonically and saturates, Whisper is U-shaped and full enhancement (α=1) falls back to the noisy baseline. The optimal α is ASR-dependent. 0.0 0.2 0.4 0.6 0.8 1.0 (phase-correction strength) 10.2 10.4 10.6 10.8 11.0 11.2 11.4 WER (%) wav2vec2.0 = 0.25 * = 0.85 = 1.0 0.0 0.2 0.4 0.6 0.8 1.0 (phase-correction strength) 5.6 5.8 6… view at source ↗
Figure 3
Figure 3. Phase sweep: WER vs. phase-correction strength γ at three fixed magnitudes. For both recognizer, reusing the noisy phase (γ=0) is optimal in almost every setting. suppression. Consequently, “over-suppression” is not a one￾size-fits-all explanation for SE-induced ASR errors. B. Phase axis To test whether the mask’s estimated phase correction helps recognition, we fix the magnitude strength and sweep γ, where γ = 0 re… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Generalization to MossFormerGAN-SE. We apply the α sweep at γ = 0 to its effective mask. The same recognizer-dependent pattern appears: wav2vec 2.0 improves monotonically, while Whisper remains U-shaped, showing that the diagnosis is not tied to a strict complex-mask m…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 10 linked inside Pith

  1. [1]

    Complex ratio masking for monaural speech separation,

    D. S. Williamson, Y . Wang, and D. Wang, “Complex ratio masking for monaural speech separation,”IEEE/ACM transactions on audio, speech, and language processing, vol. 24, no. 3, pp. 483–492, 2015

    2015

  2. [2]

    Complex spectral mapping for single-and multi-channel speech enhancement and robust asr,

    Z.-Q. Wang, P. Wang, and D. Wang, “Complex spectral mapping for single-and multi-channel speech enhancement and robust asr,” IEEE/ACM transactions on audio, speech, and language processing, vol. 28, pp. 1778–1787, 2020

    2020

  3. [3]

    End-to-end speech separation with unfolded iterative phase reconstruction,

    Z.-Q. Wang, J. Le Roux, D. Wang, and J. R. Hershey, “End-to-end speech separation with unfolded iterative phase reconstruction,”arXiv preprint arXiv:1804.10204, 2018

    arXiv 2018

  4. [4]

    A wavenet for speech denoising,

    D. Rethage, J. Pons, and X. Serra, “A wavenet for speech denoising,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5069–5073

    2018

  5. [5]

    Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,

    Y . Luo and N. Mesgarani, “Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,”IEEE/ACM trans- actions on audio, speech, and language processing, vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  6. [6]

    Metricgan+: An improved version of metricgan for speech enhancement,

    S.-W. Fu, C. Yu, T.-A. Hsieh, P. Plantinga, M. Ravanelli, X. Lu, and Y . Tsao, “Metricgan+: An improved version of metricgan for speech enhancement,”arXiv preprint arXiv:2104.03538, 2021

    arXiv 2021

  7. [7]

    Usefulness of phase spectrum in human speech perception

    K. K. Paliwal and L. D. Alsteris, “Usefulness of phase spectrum in human speech perception.” inINTERSPEECH, 2003, pp. 2117–2120

    2003

  8. [8]

    Phase processing for single-channel speech enhancement: History and recent advances,

    T. Gerkmann, M. Krawczyk-Becker, and J. Le Roux, “Phase processing for single-channel speech enhancement: History and recent advances,” IEEE signal processing Magazine, vol. 32, no. 2, pp. 55–66, 2015

    2015

Show all 43 references
  1. [9]

    How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,

    K. Iwamoto, T. Ochiai, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,”arXiv preprint arXiv:2201.06685, 2022

    2022 arXiv

  2. [10]

    When de-noising hurts: A systematic study of speech enhancement effects on modern medical asr systems,

    S. Chondhekar, V . Murukuri, R. Vasani, S. Goyal, R. Badami, A. Rana, S. SN, K. Pandia, S. Katiyar, N. Jagadeeshet al., “When de-noising hurts: A systematic study of speech enhancement effects on modern medical asr systems,”arXiv preprint arXiv:2512.17562, 2025

    2025

  3. [11]

    When denoising hinders: Revisiting zero-shot asr with sam-audio and whisper,

    A. Islam, R. Nahar, and M. E. Hamid, “When denoising hinders: Revisiting zero-shot asr with sam-audio and whisper,”arXiv preprint arXiv:2603.04710, 2026

    2026

  4. [12]

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

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

    2020

  5. [13]

    Robust speech recognition via large-scale weak supervi- sion,

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

    2023

  6. [14]

    How does end-to-end speech recognition training impact speech enhancement artifacts?

    K. Iwamoto, T. Ochiai, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “How does end-to-end speech recognition training impact speech enhancement artifacts?” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEE...

    2024

  7. [15]

    Rethinking processing distortions: Disentangling the impact of speech enhancement errors on speech recognition performance,

    T. Ochiai, K. Iwamoto, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “Rethinking processing distortions: Disentangling the impact of speech enhancement errors on speech recognition performance,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, v...

    2024

  8. [16]

    Effect of noise suppression losses on speech distortion and asr performance,

    S. Braun and H. Gamper, “Effect of noise suppression losses on speech distortion and asr performance,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 996–1000

    2022

  9. [17]

    On the compensation between magnitude and phase in speech separation,

    Z.-Q. Wang, G. Wichern, and J. Le Roux, “On the compensation between magnitude and phase in speech separation,”IEEE Signal Processing Letters, vol. 28, pp. 2018–2022, 2021

    2018

  10. [18]

    Joint noise adaptive training for robust automatic speech recognition,

    A. Narayanan and D. Wang, “Joint noise adaptive training for robust automatic speech recognition,” in2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2014, pp. 2504–2508

    2014

  11. [19]

    Learning to enhance or not: Neural network-based switching of en- hanced and observed signals for overlapping speech recognition,

    H. Sato, T. Ochiai, M. Delcroix, K. Kinoshita, N. Kamo, and T. Moriya, “Learning to enhance or not: Neural network-based switching of en- hanced and observed signals for overlapping speech recognition,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and...

    2022

  12. [20]

    Frcrn: Boosting feature representation using frequency recurrence for monaural speech enhancement,

    S. Zhao, B. Ma, K. N. Watcharasupat, and W.-S. Gan, “Frcrn: Boosting feature representation using frequency recurrence for monaural speech enhancement,” inICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2022, pp. 9281–9285

    2022

  13. [21]

    Phase- sensitive and recognition-boosted speech separation using deep recurrent neural networks,

    H. Erdogan, J. R. Hershey, S. Watanabe, and J. Le Roux, “Phase- sensitive and recognition-boosted speech separation using deep recurrent neural networks,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015, pp. 708–712

    2015

  14. [22]

    Task- aware warping factors in mask-based speech enhancement,

    Q. Wang, K. A. Lee, T. Koshinaka, K. Okabe, and H. Yamamoto, “Task- aware warping factors in mask-based speech enhancement,” in2021 29th European Signal Processing Conference (EUSIPCO). IEEE, 2021, pp. 476–480

    2021

  15. [23]

    Mask scalar prediction for improving robust automatic speech recog- nition,

    A. Narayanan, J. Walker, S. Panchapagesan, N. Howard, and Y . Koizumi, “Mask scalar prediction for improving robust automatic speech recog- nition,”arXiv preprint arXiv:2204.12092, 2022

    2022 arXiv

  16. [24]

    The third ‘chime’speech separation and recognition challenge: Dataset, task and baselines,

    J. Barker, R. Marxer, E. Vincent, and S. Watanabe, “The third ‘chime’speech separation and recognition challenge: Dataset, task and baselines,” in2015 IEEE workshop on automatic speech recognition and understanding (ASRU). IEEE, 2015, pp. 504–511

    2015

  17. [25]

    Impact of residual noise and artifacts in speech enhancement errors on intelligibility of human and machine

    S. Araki, A. Yamamoto, T. Ochiai, K. Arai, A. Ogawa, T. Nakatani, and T. Irino, “Impact of residual noise and artifacts in speech enhancement errors on intelligibility of human and machine.” inINTERSPEECH, 2023, pp. 2503–2507

    2023

  18. [26]

    Too good to be true: A study on modern automatic speech recognition for the evaluation of speech enhancement,

    D. de Oliveira, T. Peer, and T. Gerkmann, “Too good to be true: A study on modern automatic speech recognition for the evaluation of speech enhancement,”arXiv preprint arXiv:2605.12107, 2026

    2026 arXiv

  19. [27]

    A joint training framework for robust auto- matic speech recognition,

    Z.-Q. Wang and D. Wang, “A joint training framework for robust auto- matic speech recognition,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 24, no. 4, pp. 796–806, 2016

    2016

  20. [28]

    Snri target training for joint speech enhancement and recognition,

    Y . Koizumi, S. Karita, A. Narayanan, S. Panchapagesan, and M. Bacchi- ani, “Snri target training for joint speech enhancement and recognition,” arXiv preprint arXiv:2111.00764, 2021

    2021 arXiv

  21. [29]

    Dual application of speech enhancement for automatic speech recognition,

    A. Pandey, C. Liu, Y . Wang, and Y . Saraf, “Dual application of speech enhancement for automatic speech recognition,” in2021 IEEE spoken language technology workshop (SLT). IEEE, 2021, pp. 223–228

    2021

  22. [30]

    Interactive feature fusion for end-to-end noise-robust speech recognition,

    Y . Hu, N. Hou, C. Chen, and E. S. Chng, “Interactive feature fusion for end-to-end noise-robust speech recognition,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2022, pp. 6292–6296

    2022

  23. [31]

    Enhanced spectral features for distortion- independent acoustic modeling

    P. Wang and D. Wang, “Enhanced spectral features for distortion- independent acoustic modeling.” inINTERSPEECH, 2019, pp. 476–480

    2019

  24. [32]

    Should we always separate?: Switching between enhanced and ob- served signals for overlapping speech recognition,

    H. Sato, T. Ochiai, M. Delcroix, K. Kinoshita, T. Moriya, and N. Kamo, “Should we always separate?: Switching between enhanced and ob- served signals for overlapping speech recognition,”arXiv preprint arXiv:2106.00949, 2021

    2021 arXiv

  25. [33]

    Training-free intelligibility-guided observation addition for noisy asr,

    H. Li, C. Liu, W. Rao, H. Shi, S. Sakti, and E. S. Chng, “Training-free intelligibility-guided observation addition for noisy asr,”arXiv preprint arXiv:2602.20967, 2026

    2026 arXiv

  26. [34]

    Mossformer2: Combining transformer and rnn-free recurrent network for enhanced time-domain monaural speech separation,

    S. Zhao, Y . Ma, C. Ni, C. Zhang, H. Wang, T. H. Nguyen, K. Zhou, J. Q. Yip, D. Ng, and B. Ma, “Mossformer2: Combining transformer and rnn-free recurrent network for enhanced time-domain monaural speech separation,” inICASSP 2024-2024 IEEE International Conference on Acoustics...

    2024

  27. [35]

    Meanflowse: One-step generative speech enhancement via conditional mean flow,

    D. Li, S. Lu, H. Pan, Z. Zhan, Q. Hong, and L. Li, “Meanflowse: One-step generative speech enhancement via conditional mean flow,” in ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 16 172–16 176

    2026

  28. [36]

    Flowse: Efficient and high-quality speech enhancement via flow matching,

    Z. Wang, Z. Liu, X. Zhu, Y . Zhu, M. Liu, J. Chen, L. Xiao, C. Weng, and L. Xie, “Flowse: Efficient and high-quality speech enhancement via flow matching,”arXiv preprint arXiv:2505.19476, 2025

    2025 arXiv

  29. [37]

    Investigating rnn-based speech enhancement methods for noise-robust text-to-speech,

    C. V . Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “Investigating rnn-based speech enhancement methods for noise-robust text-to-speech,” in9th ISCA speech synthesis workshop, 2016, pp. 159–165

    2016

  30. [38]

    The diverse environments multi- channel acoustic noise database (demand): A database of multichannel environmental noise recordings,

    J. Thiemann, N. Ito, and E. Vincent, “The diverse environments multi- channel acoustic noise database (demand): A database of multichannel environmental noise recordings,” inProceedings of Meetings on Acous- tics, vol. 19, no. 1. Acoustical Society of America, 2013, p. 035081

    2013

  31. [39]

    Clearervoice-studio: Bridging advanced speech processing research and practical deployment,

    S. Zhao, Z. Pan, and B. Ma, “Clearervoice-studio: Bridging advanced speech processing research and practical deployment,”arXiv preprint arXiv:2506.19398, 2025

    2025 arXiv

  32. [40]

    Performance measurement in blind audio source separation,

    E. Vincent, R. Gribonval, and C. F ´evotte, “Performance measurement in blind audio source separation,”IEEE transactions on audio, speech, and language processing, vol. 14, no. 4, pp. 1462–1469, 2006

    2006

  33. [41]

    Per- ceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Per- ceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,” in2001 IEEE international conference on acoustics, speech, and signal processing. Procee...

    2001

  34. [42]

    An algorithm for intelligibility prediction of time–frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An algorithm for intelligibility prediction of time–frequency weighted noisy speech,” IEEE Transactions on audio, speech, and language processing, vol. 19, no. 7, pp. 2125–2136, 2011

    2011

  35. [43]

    Bridging the gap: Integrating pre-trained speech enhancement and recognition models for robust speech recognition,

    K.-C. Wang, Y .-J. Li, W.-L. Chen, Y .-W. Chen, Y .-C. Wang, P.-C. Yeh, C. Zhang, and Y . Tsao, “Bridging the gap: Integrating pre-trained speech enhancement and recognition models for robust speech recognition,” in 2024 32nd European Signal Processing Conference (EUSIPCO). IE...

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.