Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Improving Speech Enhancement with Multi-Metric Supervision from Learned Quality Assessment

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A frozen multi-metric quality assessment model can supervise speech enhancement training, improving perceptual metrics on simulated and real-world audio.

desk verdict Solid SQA-supervised SE recipe, but the real-data evidence for consistent gains is thinner than the abstract claims. read the letter →

arxiv 2506.12260 v2 pith:4VGRVYEB submitted 2025-06-13 cs.SD eess.AS

Wei Wang , Wangyou Zhang , Chenda Li , Jiatong Shi , Shinji Watanabe , Yanmin Qian This is my paper
classification cs.SDeess.AS
keywords speechenhancementqualityassessmentmulti-metricsupervisiondifferentiablelossreal-worldtrainingadversarialcollapseself-supervisedregularizationUni-VERSA-Ext
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

This paper argues that a learned multi-metric speech quality assessment (SQA) model can act as a training signal for speech enhancement, replacing or supplementing conventional objectives such as SI-SNR that correlate poorly with perception and require clean reference signals. The authors extend the Uni-VERSA model into Uni-VERSA-Ext, which predicts 22 quality metrics from a single utterance, and use it as a frozen, differentiable supervisor while fine-tuning a BSRNN enhancement network. On simulated data the training objective combines a spectrogram loss, a score-based loss, and a feature-space loss; on real data without references, it combines the score loss with a self-supervised regularization that keeps enhanced output near the initial model's output. The paper reports consistent improvements across perceptual metrics on DNS-2020 and CHiME-4 test sets, and shows that score-only supervision collapses into adversarial non-speech outputs that the SQA model wrongly scores as high quality. The reason to care is that this points to a way of optimizing enhancement systems directly for perceived quality, and of using unpaired real-world recordings in training.

What carries the argument

The load-bearing object is Uni-VERSA-Ext, a multi-metric speech quality assessment model that shares the Uni-VERSA architecture but predicts 22 quality metrics and exposes a penultimate-layer hidden feature vector. Trained on over 500k enhanced utterances from the URGENT 2024 and 2025 challenges, it is frozen during enhancement fine-tuning and used in two ways: the score loss $\mathcal{L}_{\mathrm{score}}$ is a sign- and weight-adjusted sum of predicted metric scores, making non-differentiable or expensive metrics such as PESQ, ESTOI, speaker similarity, and CER optimizable; the feature loss $\mathcal{L}_{\mathrm{feat}}$ is the L1 distance between hidden features of enhanced and clean speech. For real-world data, the regularization term $\mathcal{L}_{\mathrm{reg}}$ measures the distance between the spectrogram of the current enhanced output and that of the initial model's output, anchoring the fine-tuned model to its pretrained behavior and preventing it from exploiting the SQA model's blind spots.

What would settle it

Run a listening test on the DNS-2020 non-blind test set comparing the SQA-fine-tuned model against the spectrogram-only baseline; if the SQA-guided model does not achieve higher human MOS while its predicted metrics improve, the claimed alignment between the proxy and perception fails. A cheaper computational check is to measure the per-utterance correlation between changes in Uni-VERSA-Ext-predicted scores and changes in human MOS; zero or negative correlation would falsify the core premise.

Watch

Extended reading notes

Core claim

The paper's central claim is that SQA-guided training, using Uni-VERSA-Ext as a frozen differentiable proxy for 22 evaluation metrics, consistently improves speech enhancement performance on both simulated and real-world test sets. Concretely, fine-tuning a pretrained enhancement model with $\mathcal{L}_{\mathrm{simu}} = \lambda_{\mathrm{spec}}\mathcal{L}_{\mathrm{spec}} + \lambda_{\mathrm{score}}\mathcal{L}_{\mathrm{score}} + \lambda_{\mathrm{feat}}\mathcal{L}_{\mathrm{feat}}$ on simulated data, and with $\mathcal{L}_{\mathrm{real}} = \lambda_{\mathrm{score}}\mathcal{L}_{\mathrm{score}} + \lambda_{\mathrm{reg}}\mathcal{L}_{\mathrm{reg}}$ on real data, improves a range of perceptual metrics, including UTMOS, DNSMOS, and the URGENT ranking score. The authors further claim that the self-supervised term $\mathcal{L}_{\mathrm{reg}}$ is necessary: score-only supervision leads to complete collapse (the model outputs non-speech signals that the SQA model misrates as high quality), and on real data without $\mathcal{L}_{\mathrm{reg}}$ the training diverges. The feature-space loss $\mathcal{L}_{\mathrm{feat}}$, which pulls the SQA features of enhanced speech toward those of clean speech, is a key contributor on simulated data.

Load-bearing premise

The load-bearing premise is that the frozen Uni-VERSA-Ext model's predicted scores and internal features are trustworthy enough that pushing them upward or closer to clean-speech features genuinely improves perceived quality, and that the proposed regularization term is sufficient to stop the enhancement network from gaming the assessor on real data.

Editorial extensions

If this is right

  • Fine-tuning a pretrained enhancement model with the combined SQA objective improves non-intrusive perceptual metrics (UTMOS, DNSMOS, SIGMOS, Distill-MOS) on both simulated and real test sets relative to spectrogram-only fine-tuning.
  • Non-differentiable and computationally heavy evaluation metrics become directly optimizable as differentiable proxies through the frozen SQA model.
  • Real-world recordings without clean references can be incorporated into enhancement training, reducing the simulation-to-real mismatch.
  • Score-only SQA supervision is unstable: without a reference-anchored loss it collapses to adversarial non-speech outputs, and without $\mathcal{L}_{\mathrm{reg}}$ real-data training diverges.
  • Scaling simulated training data mitigates some side effects, such as CER degradation on non-reverberant test sets.

Reading between the lines

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

  • The same frozen-assessor-plus-anchor recipe should transfer to other generative audio tasks (text-to-speech, voice conversion, music enhancement) where a learned quality model exists and paired references are scarce.
  • The documented collapse suggests a broader design rule for using any differentiable proxy as a training loss: a reference-anchored or distribution-anchored term is not an add-on but a necessary stabilizer, and $\mathcal{L}_{\mathrm{reg}}$ is one instance of that family.
  • The paper does not sweep the regularization weight $\lambda_{\mathrm{reg}}$; testing whether the margin before divergence is sensitive to that weight, and to the choice of backbone, would tell how general the stabilizer is.
  • Because Uni-VERSA-Ext is trained on challenge submissions, its metric coverage may be biased toward the distortion types in those corpora; out-of-domain generalization of the supervised gains (e.g., to unseen noise types or languages) remains an open question.
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

4 major / 4 minor

Summary. The paper proposes a framework for using a learned multi-metric speech quality assessment (SQA) model, Uni-VERSA-Ext, as a supervisory signal for speech enhancement (SE) training. The SQA model is extended from Uni-VERSA by expanding training data, the metric space (from 11 to 22 metrics), and output activation functions. The SE model (BSRNN) is fine-tuned with a combination of spectrogram loss, score-based SQA loss, and feature-based SQA loss on simulated data, and with score-based loss plus a self-supervised regularization on real data to prevent adversarial exploitation. Experiments on DNS-2020 and CHiME-4 ET05 report improvements on several non-intrusive quality metrics and CER, but the real-data improvements are small and mixed, and a key table contains a duplicated row.

Significance. The paper addresses an important problem: aligning SE training with perceptual quality and enabling training on real-world data without clean references. The introduction of Uni-VERSA-Ext, which achieves higher correlation with human MOS on in-domain and out-of-domain SQA benchmarks (Tables III and IV), is a solid contribution. The paper also provides code and checkpoints, which supports reproducibility. However, the central claim that SQA-guided training consistently improves SE performance is not convincingly established. The simulated-data results show the score loss adds little beyond feature and spectrogram losses, and the real-data results are dominated by tiny, mixed differences without statistical assessment. The evaluation metrics overlap with supervision targets, creating a reward-hacking risk that is not addressed with an independent perceptual evaluation. The duplicate rows in Table VII further weaken the evidence. The paper's contributions are useful but require revision to substantiate the claims.

major comments (4)
  1. [§III-E, Table VI] The real-data results do not support the claim of 'consistent improvements' in the abstract. On the CHiME-4 ET05 test set, the SQA-guided fine-tuning (λscore/λreg = 1/1) versus baseline shows: DNSMOS decreases from 2.71 to 2.70, UTMOS increases from 2.65 to 2.66, NISQA increases from 3.02 to 3.06, SIGMOS.OVRL is unchanged at 2.38, Distill-MOS decreases from 2.60 to 2.59, and only CER improves from 8.13 to 7.95. These differences are at or below 0.04 units, and no error bars, significance tests, or multiple-seed variance are reported. Such small and mixed changes do not establish a consistent improvement.
  2. [§III-E, Table VII] Table VII contains an identical row for the 100h and 2500h data-scale settings (DNS. 2.97, UT. 3.35, NISQA 4.18, SIG.OVRL 2.85, Distill. 3.60, CER 7.79), indicating a copy-paste error. This table is the primary evidence for the benefit of SQA supervision with mixed simulated/real data. Moreover, at 700h, the SQA-guided model worsens CER from 6.88 to 7.81 while improving other metrics, which contradicts the 'consistent improvements' claim. The duplicated row and the CER regression must be explained or corrected.
  3. [§II-E2, Eq. (6) and §III-D, Table V] The proposed supervision suffers from a significant reward-hacking risk that is not adequately mitigated or evaluated. Training with Lscore alone causes collapse to non-speech (Table V, line 8: SDR 1.56/-1.43, PESQ 1.04/1.03), and on real data without Lreg, training diverges (Table VI: '1/0 (diverge)'). The regularization Lreg (Eq. 6) stabilizes training, but the evaluation metrics (DNSMOS, UTMOS, NISQA, Distill-MOS) are from the same family as the supervision targets, so improvements on these metrics could reflect overfitting to the SQA model rather than genuine perceptual improvement. The paper does not report a human listening test or an independent intrusive metric (e.g., PESQ, SDR with clean references) on real data, leaving the perceptual validity of the gains unverified.
  4. [§III-D, Table V] The contribution of the score-based loss on simulated data is marginal. Comparing line 4 (λspec/λscore/λfeat = 1/1/1) with line 6 (1/0/1) shows nearly identical results across all metrics, and line 7 (0/1/1) versus line 9 (0/0/1) are also nearly indistinguishable. This suggests that the score-based SQA supervision provides little benefit beyond the feature loss and spectrogram loss on simulated data. The paper should either demonstrate a scenario where Lscore is essential (beyond avoiding the collapse case) or temper the claim that SQA-guided training is the key driver of improvement.
minor comments (4)
  1. [Table VII] The duplicated rows for 100h and 2500h must be corrected; if the 2500h results are actually the same as the 100h results, the authors should explain why, but a verbatim duplication is likely an error.
  2. [§III-B] The values of the loss weights λspec, λscore, λfeat, and λreg are not specified in the text. Please provide the hyperparameters used in the experiments, as they are central to the proposed objective.
  3. [Eq. (3)] In Eq. (3), the notation uses h(˜xsimu) and h(ˆxsimu) for enhanced and clean signals, but the text refers to 'enhanced and clean reference signals.' Clarify the notation to avoid ambiguity between the hat and tilde symbols.
  4. [§III-D, Table V] The paper reports results as (Synthetic / Synthetic Reverb) but does not provide confidence intervals or variance across model seeds; given the small differences between configurations, some measure of variability (e.g., multiple fine-tuning runs) would strengthen the conclusions.

Circularity Check

1 steps flagged · score 4.0 of 10

SQA-model benchmark is circular due to train/test overlap; main SE claim is surrogate optimization, not construction-level circularity.

  1. fitted input called prediction [Section III-A (Table II note) and Section III-C]
    "Although the URGENT challenge datasets refer to these sets as 'test' sets (e.g., blind/non-blind test), they are repurposed as training resources... For in-domain evaluation, we follow [23] and use a subset of the URGENT 2024 non-blind test set, which comprises 4,200 enhanced utterances derived from 50 source recordings."

    Table II lists 'UG24 Non-blind Test 153,900 850 285.16' among the Uni-VERSA-Ext training clips, and the in-domain benchmark is drawn from the same UG24 non-blind test set. Thus the improved LCC/SRCC reported in Table III are in-sample fits rather than predictive generalization: the model was trained on the very set from which the 'evaluation' subset is taken. The paper even acknowledges these sets are 'repurposed as training resources,' so presenting the resulting correlations as evidence that the extended model 'better captures perceptual quality' is a fitted-input result, not an independent prediction.

full rationale

The paper's central SE derivation is not circular by construction: Lspec, Lfeat, and Lreg are independent signal/feature/regularization terms, and the SE evaluations on DNS-2020 and CHiME-4 use held-out test sets with the official metric implementations. The score-based loss is a learned differentiable surrogate of the evaluation metrics, which is the intended mechanism, and the paper explicitly documents the adversarial failure mode of that surrogate (Table V line 8, Figure 2) and introduces Lreg to counter it. However, one load-bearing component is circular: Uni-VERSA-Ext is trained on the URGENT 2024 non-blind test set (153,900 clips, Table II) and then evaluated 'in-domain' on a subset of that same non-blind test set (Section III-A), so the Table III claim of improved SQA correlation is an in-sample fit, not a prediction. This does not by itself invalidate the SE results, but it undermines the independent demonstration of the extended SQA model's generalization. The real-data SE evidence is additionally weak—tiny deltas, mixed CER, and identical 100h/2500h rows in Table VII—but those are evidence-quality concerns rather than construction-level circularity. Overlapping-author citations such as [23] and [39] are used for architecture and for the ranking metric's motivation, but they are not load-bearing in a way that forces the central SE conclusion.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The method relies on a trained SQA model, Uni-VERSA-Ext, which is itself an extension of prior work by overlapping authors. The main free choices are the loss weighting scheme and the regularization term, both justified only empirically.

free parameters (2)
  • loss weights = lambda_spec=1, lambda_score=1, lambda_feat=1 (simulated); lambda_score=1, lambda_reg=1 (real)
    Chosen by hand with no sensitivity analysis. The central training objectives in Eqs. (5) and (7) depend directly on these weights.
  • metric weights w_k = 1 for all metrics
    Default to 1 unless otherwise specified. No tuning procedure is described, and the score loss in Eq. (2) sums over these weights.
assumptions (3)
  • domain assumption Predicted SQA scores are differentiable proxies for the target metrics, and optimizing them improves the actual metrics.
    Invoked in Section II-C, Eq. (2). If false, the approach reduces to reward hacking, as shown by the adversarial collapse in Table V line 8.
  • domain assumption L1 distance in SQA hidden feature space between enhanced and clean speech provides useful supervision.
    Invoked in Section II-D, Eq. (3). No analysis is given of what these features encode or why L1 distance in this space is perceptually meaningful.
  • domain assumption Lreg keeps the fine-tuned model close to the initial model and prevents out-of-distribution outputs.
    Invoked in Section II-E2, Eq. (6). The weight lambda_reg=1 is chosen without sensitivity analysis, and the mechanism by which it prevents adversarial exploitation is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Speech Enhancement with Multi-Metric Supervision from Learned Quality Assessment." pith.science (2026). https://pith.science/paper/4VGRVYEB

@misc{pith2026250612260,
  author       = {Pith},
  title        = {Pith review of: Improving Speech Enhancement with Multi-Metric Supervision from Learned Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VGRVYEB}},
  note         = {Machine review of arXiv:2506.12260}
}
read the original abstract

Speech quality assessment (SQA) aims to predict the perceived quality of speech signals under a wide range of distortions. It is inherently connected to speech enhancement (SE), which seeks to improve speech quality by removing unwanted signal components. While SQA models are widely used to evaluate SE performance, their potential to guide SE training remains underexplored. In this work, we investigate a training framework that leverages a SQA model, trained to predict multiple evaluation metrics from a public SE leaderboard, as a supervisory signal for SE. This approach addresses a key limitation of conventional SE objectives, such as SI-SNR, which often fail to align with perceptual quality and generalize poorly across evaluation metrics. Moreover, it enables training on real-world data where clean references are unavailable. Experiments on both simulated and real-world test sets show that SQA-guided training consistently improves performance across a range of quality metrics. Code and checkpoints are available at https://github.com/urgent-challenge/urgent2026_challenge_track2

Figures

Figures reproduced from arXiv: 2506.12260 by the authors.

Figure 1
Figure 1. Overview of the training pipeline for SQA-guided SE using either [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An example of an adversarial sample resulting from training with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Less is More: Data Curation Matters in Scaling Speech Enhancement

    eess.AS 2025-06 conditional novelty 5.0 of 10

    A quality-filtered 700-hour subset beats the full 2,500-hour URGENT2025 training set on perceptual quality metrics for both discriminative and generative speech enhancement models.

Reference graph

Works this paper leans on

54 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    P. C. Loizou, Speech enhancement: theory and practice . CRC press, 2013

    2013

  2. [2]

    SDR–half- baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR–half- baked or well done?” in ICASSP, 2019, pp. 626–630

    work page 2019

  3. [3]

    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,” in Proc. Interspeech 2022 , 2022, pp. 5418– 5422

    work page 2022

  4. [4]

    Bridging the gap between monaural speech enhancement and recognition with distortion-independent acous- tic modeling,

    P. Wang, K. Tan, and D. L. Wang, “Bridging the gap between monaural speech enhancement and recognition with distortion-independent acous- tic modeling,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 39–48, 2020

    work page 2020

  5. [5]

    Advancing non-intrusive suppression on enhancement distortion for noise robust asr,

    W. Wang, S. Zhao, and Y . Qian, “Advancing non-intrusive suppression on enhancement distortion for noise robust asr,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    work page 2025

  6. [6]

    Fat-hubert: Front-end adaptive training of hidden-unit bert for distortion-invariant robust speech recognition,

    D. Yang, W. Wang, and Y . Qian, “Fat-hubert: Front-end adaptive training of hidden-unit bert for distortion-invariant robust speech recognition,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    work page 2023

  7. [7]

    Closing the gap between time-domain multi-channel speech enhancement on real and simulation conditions,

    W. Zhang, J. Shi, C. Li, S. Watanabe, and Y . Qian, “Closing the gap between time-domain multi-channel speech enhancement on real and simulation conditions,” in 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) . IEEE, 2021, pp. 146–150

    work page 2021

  8. [8]

    Less is more: Data curation matters in scaling speech enhancement,

    C. Li, W. Zhang, W. Wang, R. Scheibler, K. Saijo, S. Cornell, Y . Fu, M. Sach, Z. Ni, A. Kumar et al., “Less is more: Data curation matters in scaling speech enhancement,” arXiv preprint arXiv:2506.23859 , 2025

    arXiv 2025

Show all 54 references
  1. [9]

    Lightweight Front-end Enhancement for Robust ASR via Frame Resampling and Sub-Band Pruning,

    Siyi Zhao and Wei Wang and Yanmin Qian, “Lightweight Front-end Enhancement for Robust ASR via Frame Resampling and Sub-Band Pruning,” in Interspeech 2025, 2025, pp. 3409–3413

    2025

  2. [10]

    A review on subjective and objective evaluation of synthetic speech,

    E. Cooper, W.-C. Huang, Y . Tsao, H.-M. Wang, T. Toda, and J. Ya- magishi, “A review on subjective and objective evaluation of synthetic speech,” Acoustical Science and Technology , vol. advpub, p. e24.12, 2024

    2024

  3. [11]

    Objective measures of perceptual audio quality reviewed: An evaluation of their application domain dependence,

    M. Torcoli, T. Kastner, and J. Herre, “Objective measures of perceptual audio quality reviewed: An evaluation of their application domain dependence,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 1530–1541, 2021

    2021

  4. [12]

    Versa: A versatile evaluation toolkit for speech, audio, and music,

    J. Shi, H.-j. Shim, J. Tian, S. Arora, H. Wu, D. Petermann, J. Q. Yip, Y . Zhang, Y . Tang, W. Zhanget al., “Versa: A versatile evaluation toolkit for speech, audio, and music,” arXiv e-prints , pp. arXiv–2412, 2024

    2024

  5. [13]

    Lessons learned from the urgent 2024 speech enhancement challenge,

    W. Zhang, K. Saijo, S. Cornell, R. Scheibler, C. Li, Z. Ni, A. Kumar, M. Sach, W. Wang, Y . Fuet al., “Lessons learned from the urgent 2024 speech enhancement challenge,” arXiv preprint arXiv:2506.01611, 2025

    2024 arXiv

  6. [14]

    Meta audiobox aesthetics: Unified automatic quality assessment for speech, music, and sound,

    A. Tjandra, Y .-C. Wu, B. Guo, J. Hoffman, B. Ellis, A. Vyas, B. Shi, S. Chen, M. Le, N. Zacharov, C. Wood, A. Lee, and W.-N. Hsu, “Meta audiobox aesthetics: Unified automatic quality assessment for speech, music, and sound,” ArXiv, vol. abs/2502.05139, 2025. [Online]. Availab...

    2025 arXiv

  7. [15]

    DNSMOS P.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. Reddy, V . Gopal, and R. Cutler, “DNSMOS P.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in ICASSP, 2022, pp. 886–890

    2022

  8. [16]

    UTMOS: UTokyo-SaruLab system for V oiceMOS chal- lenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: UTokyo-SaruLab system for V oiceMOS chal- lenge 2022,” in Interspeech, 2022, pp. 4521–4525

    2022

  9. [17]

    The t05 system for the voicemos challenge 2024: Transfer learning from deep image classifier to naturalness mos prediction of high-quality synthetic speech,

    K. Baba, W. Nakata, Y . Saito, and H. Saruwatari, “The t05 system for the voicemos challenge 2024: Transfer learning from deep image classifier to naturalness mos prediction of high-quality synthetic speech,” in IEEE Spoken Language Technology Workshop (SLT) . IEEE, 2024, pp. 818– 824

    2024

  10. [18]

    The voicemos challenge 2022,

    W. C. Huang, E. Cooper, Y . Tsao, H.-M. Wang, T. Toda, and J. Yamag- ishi, “The voicemos challenge 2022,” in Proc. Interspeech 2022 , 2022, pp. 4536–4540

    2022

  11. [19]

    The voicemos challenge 2023: Zero-shot subjective speech quality prediction for multiple domains,

    E. Cooper, W.-C. Huang, Y . Tsao, H.-M. Wang, T. Toda, and J. Yamag- ishi, “The voicemos challenge 2023: Zero-shot subjective speech quality prediction for multiple domains,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . IEEE, 2023, pp. 1–7

    2023

  12. [20]

    The voicemos challenge 2024: Beyond speech quality prediction,

    W.-C. Huang, S.-W. Fu, E. Cooper, R. E. Zezario, T. Toda, H.-M. Wang, J. Yamagishi, and Y . Tsao, “The voicemos challenge 2024: Beyond speech quality prediction,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 803–810

    2024

  13. [21]

    Urgent-pk: Perceptually-aligned ranking model designed for speech enhancement competition,

    J. Wang, C. Li, W. Wang, W. Zhang, S. Cornell, M. Sach, R. Scheibler, K. Saijo, Y . Fu, Z. Ni et al. , “Urgent-pk: Perceptually-aligned ranking model designed for speech enhancement competition,” arXiv preprint arXiv:2506.23874, 2025

    2025 arXiv

  14. [22]

    ICASSP 2024 speech signal improvement challenge,

    N.-C. Ristea, B. Naderi, A. Saabas, R. Cutler, S. Braun, and S. Branets, “ICASSP 2024 speech signal improvement challenge,” IEEE Open Journal of Signal Processing , vol. 6, pp. 238–246, 2025

    2024

  15. [23]

    Uni-versa: Versatile speech assessment with a unified network,

    J. Shi, H.-J. Shim, and S. Watanabe, “Uni-versa: Versatile speech assessment with a unified network,” arXiv preprint arXiv:2505.20741 , 2025

    2025 arXiv

  16. [24]

    Perceptual 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, “Perceptual evaluation of speech quality (PESQ)—a new method for speech quality assessment of telephone networks and codecs,” in ICASSP, 2001, pp. 749–752

    2001

  17. [25]

    Perceptual objective listening quality assess- ment (POLQA), the third generation ITU-T standard for end-to-end speech quality measurement part I–—temporal alignment,

    J. G. Beerends, C. Schmidmer, J. Berger, M. Obermann, R. Ullmann, J. Pomy, and M. Keyhl, “Perceptual objective listening quality assess- ment (POLQA), the third generation ITU-T standard for end-to-end speech quality measurement part I–—temporal alignment,” AES Journal, vol. 6...

    2013

  18. [26]

    URGENT challenge: Universality, robustness, and generalizability for speech en- hancement,

    W. Zhang, R. Scheibler, K. Saijo, S. Cornell, C. Li, Z. Ni, J. Pirkl- bauer, M. Sach, S. Watanabe, T. Fingscheidt, and Y . Qian, “URGENT challenge: Universality, robustness, and generalizability for speech en- hancement,” in Interspeech, 2024, pp. 4868–4872

    2024

  19. [27]

    Inter- speech 2025 URGENT speech enhancement challenge,

    K. Saijo, W. Zhang, S. Cornell, R. Scheibler, C. Li, Z. Ni, A. Kumar, M. Sach, Y . Fu, W. Wang, T. Fingscheidt, and S. Watanabe, “Inter- speech 2025 URGENT speech enhancement challenge,” Accepted by Interspeech, 2025

    2025

  20. [28]

    Performance measurement in blind audio source separation,

    E. Vincent, R. Gribonval, and C. Févotte, “Performance measurement in blind audio source separation,” IEEE Trans. ASLP ., vol. 14, no. 4, pp. 1462–1469, 2006

    2006

  21. [29]

    Distillation and pruning for scalable self- supervised representation-based speech quality assessment,

    B. Stahl and H. Gamper, “Distillation and pruning for scalable self- supervised representation-based speech quality assessment,” in ICASSP, 2025

    2025

  22. [30]

    NISQA: A deep CNN- self-attention model for multidimensional speech quality prediction with crowdsourced datasets,

    G. Mittag, B. Naderi, A. Chehadi, and S. Möller, “NISQA: A deep CNN- self-attention model for multidimensional speech quality prediction with crowdsourced datasets,” in Interspeech, 2021, pp. 2127–2131

    2021

  23. [31]

    SCOREQ: Speech quality assessment with contrastive regression,

    A. Ragano, J. Skoglund, and A. Hines, “SCOREQ: Speech quality assessment with contrastive regression,” in Advances in Neural Infor- mation Processing Systems , vol. 37, 2024, pp. 105 702–105 729

    2024

  24. [32]

    Owsm v3. 1: Better and faster open whisper-style speech models based on e-branchformer,

    Y . Peng, J. Tian, W. Chen, S. Arora, B. Yan, Y . Sudo, M. Shakeel, K. Choi, J. Shi, X. Chang et al. , “Owsm v3. 1: Better and faster open whisper-style speech models based on e-branchformer,” in Proc. Interspeech 2024 , 2024, pp. 352–356

    2024

  25. [33]

    An algorithm for predicting the intelligibility of speech masked by modulated noise maskers,

    J. Jensen and C. H. Taal, “An algorithm for predicting the intelligibility of speech masked by modulated noise maskers,” IEEE/ACM Trans. ASLP ., vol. 24, no. 11, pp. 2009–2022, 2016

    2009

  26. [34]

    SpeechBERTScore: Reference-aware automatic evaluation of speech generation leveraging NLP evaluation metrics,

    T. Saeki, S. Maiti, S. Takamichi, S. Watanabe, and H. Saruwatari, “SpeechBERTScore: Reference-aware automatic evaluation of speech generation leveraging NLP evaluation metrics,” in Interspeech, 2024, pp. 4943–4947

    2024

  27. [35]

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

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language processing , vol. 29, pp. 3451–3460, 2021

    2021

  28. [36]

    Evaluation metrics for generative speech enhancement methods: Issues and perspectives,

    J. Pirklbauer, M. Sach, K. Fluyt, W. Tirry, W. Wardah, S. Moeller, and T. Fingscheidt, “Evaluation metrics for generative speech enhancement methods: Issues and perspectives,” in Speech Communication; 15th ITG Conference, 2023, pp. 265–269

    2023

  29. [37]

    Espnet-spk: full pipeline speaker embedding toolkit with reproducible recipes, self- supervised front-ends, and off-the-shelf models,

    J.-w. Jung, W. Zhang, J. Shi, Z. Aldeneh, T. Higuchi, A. Gichamba, B.-J. Theobald, A. Hussen Abdelaziz, and S. Watanabe, “Espnet-spk: full pipeline speaker embedding toolkit with reproducible recipes, self- supervised front-ends, and off-the-shelf models,” in Proc. Interspeech...

    2024

  30. [38]

    Mel-cepstral distance measure for objective speech qual- ity assessment,

    R. Kubichek, “Mel-cepstral distance measure for objective speech qual- ity assessment,” in Proc. PACRIM, 1993, pp. 125–128

    1993

  31. [39]

    Lessons learned from the URGENT 2024 speech enhancement chal- lenge,

    W. Zhang, K. Saijo, S. Cornell, R. Scheibler, C. Li, Z. Ni, A. Kumar, M. Sach, W. Wang, Y . Fu, S. Watanabe, T. Fingscheidt, and Y . Qian, “Lessons learned from the URGENT 2024 speech enhancement chal- lenge,” Accepted by Interspeech , 2025

    2024

  32. [40]

    Distance measures for speech processing,

    A. Gray and J. Markel, “Distance measures for speech processing,” IEEE Transactions on Acoustics, Speech, and Signal Processing , vol. 24, no. 5, pp. 380–391, 1976

    1976

  33. [41]

    SHEET: A multi-purpose open-source speech human evaluation estimation toolkit,

    W.-C. Huang, E. Cooper, and T. Toda, “SHEET: A multi-purpose open-source speech human evaluation estimation toolkit,” arXiv preprint arXiv:2505.15061, 2025

    2025 arXiv

  34. [42]

    The chime-7 udase task: Unsupervised domain adaptation for conversational speech enhancement,

    S. Leglaive, L. Borne, E. Tzinis, M. Sadeghi, M. Fraticelli, S. Wis- dom, M. Pariente, D. Pressnitzer, and J. R. Hershey, “The chime-7 udase task: Unsupervised domain adaptation for conversational speech enhancement,” in 7th International Workshop on Speech Processing in Every...

    2023

  35. [43]

    Generalization ability of mos prediction networks,

    E. Cooper, W.-C. Huang, T. Toda, and J. Yamagishi, “Generalization ability of mos prediction networks,” in ICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8442–8446

    2022

  36. [44]

    The blizzard challenge 2019,

    Z. Wu, Z. Xie, and S. King, “The blizzard challenge 2019,” in The Blizzard Challenge 2019 , 2019, pp. 1–24

    2019

  37. [45]

    Mos-bench: Benchmarking generalization abilities of subjective speech quality assessment models,

    W.-C. Huang, E. Cooper, and T. Toda, “Mos-bench: Benchmarking generalization abilities of subjective speech quality assessment models,”

  38. [46]

    The INTERSPEECH 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,

    C. K. Reddy, V . Gopal, R. Cutler, E. Beyrami, R. Cheng, H. Dubey, S. Matusevych, R. Aichner, A. Aazami, S. Braun, P. Rana, S. Srinivasan, and J. Gehrke, “The INTERSPEECH 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” in ...

    2020

  39. [47]

    An analysis of environment, microphone and data simulation mismatches in robust speech recognition,

    E. Vincent, S. Watanabe, A. A. Nugraha, J. Barker, and R. Marxer, “An analysis of environment, microphone and data simulation mismatches in robust speech recognition,” Computer Speech & Language , vol. 46, pp. 535–557, 2017

    2017

  40. [48]

    Espnet: End-to-end speech processing toolkit,

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. E. Y . Soplin, J. Heymann, M. Wiesner, N. Chen et al. , “Espnet: End-to-end speech processing toolkit,” arXiv preprint arXiv:1804.00015, 2018

    2018 arXiv

  41. [49]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao et al. , “Wavlm: Large-scale self-supervised pre- training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  42. [50]

    SUPERB: Speech Processing Universal PERformance Benchmark,

    S. wen Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K. tik Lee, D.-R. Liu, Z. Huang, S. Dong, S.-W. Li, S. Watanabe, A. Mohamed, and H. yi Lee, “SUPERB: Speech Processing Universal PER...

    2021

  43. [51]

    Music source separation with band-split rnn,

    Y . Luo and J. Yu, “Music source separation with band-split rnn,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 1893–1901, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:252668641

    1901

  44. [52]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,” in International Conference on Learning Representations , 2018

    2018

  45. [53]

    Adversarial attacks on automatic speech recognition (asr): A survey,

    A. R. Bhanushali, H. Mun, and J. Yun, “Adversarial attacks on automatic speech recognition (asr): A survey,” IEEE Access , 2024

    2024

  46. [2024]

    Available: https://arxiv.org/abs/2411.03715

    [Online]. Available: https://arxiv.org/abs/2411.03715

    arXiv

Pith tools

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