Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Less is More: Data Curation Matters in Scaling Speech Enhancement

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Models trained on a curated 700-hour subset of speech enhancement data outperform models trained on the full 2,500-hour set on every reported non-intrusive quality metric.

desk verdict A useful, well-run study showing curation beats bulk data in speech enhancement, but the headline causal claim is undercut by an uncontrolled composition shift. read the letter →

arxiv 2506.23859 v2 pith:OETW6E43 submitted 2025-06-30 eess.AS cs.SD

Chenda Li , Wangyou Zhang , Wei Wang , Robin Scheibler , Kohei Saijo , Samuele Cornell , Yihui Fu , Marvin Sach
show 5 more authors
Zhaoheng Ni Anurag Kumar Tim Fingscheidt Shinji Watanabe Yanmin Qian
This is my paper
classification eess.AScs.SD
keywords speechenhancementdatacurationthreshold-basedfilteringnon-intrusivequalitymetricsscalinglawsBSRNNflowmatchingURGENT2025
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that in modern speech enhancement, the usual assumption that more training data is always better fails once the data contains low-quality "clean" labels. Using the 2,500-hour URGENT2025 training set, it applies threshold-based filtering with five non-intrusive quality metrics and keeps roughly 700 hours of high-quality speech. Models trained on this curated subset outperform models trained on the full dataset on all six reported non-intrusive metrics, for both a discriminative BSRNN and a generative flow-matching model, with intrusive metrics on par or slightly worse. The paper concludes that data curation, not raw data volume, is the lever that matters when scaling speech enhancement systems.

What carries the argument

The load-bearing mechanism is the two-stage curation pipeline: threshold-based filtering (TBF) scores each "clean" utterance with five non-intrusive DNN quality metrics (DNSMOS, NISQA, SIGMOS, Torchaudio-SQUIM-SDR, UTMOS), each source dataset getting its own threshold, and keeps only utterances that clear every threshold; then a quality-ranking stage normalizes the metric scores per utterance, sums them, and selects top-100, top-350, and top-700-hour slices. This pipeline is what converts the hypothesis that dirty "clean" labels hurt scaling into an operational selection rule, and it is what the controlled comparison against uniform random subsets rests on.

What would settle it

Train a control model on a 700-hour subset that matches the full set's per-dataset composition but keeps only utterances passing the same quality thresholds; if that control reproduces the 700h-TBF results, quality filtering is the cause, and if not, the gains come from changed composition.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that "less is more" holds for universal speech enhancement: threshold-based filtering of the 2,500-hour URGENT2025 Track1 training data down to about 700 hours yields models that beat the full-data models on every non-intrusive metric in the evaluation, including DNSMOS Pro and DistillMOS, which were not part of the filtering. The same pattern holds for both a discriminative band-split RNN and a generative flow-matching model, and quality-ranked subsets outperform uniform random subsets at every size tested. The paper also finds that initializing from the full-data checkpoint and then training on the curated subset gives the best of both worlds, with the largest gains on perceptual-quality metrics and no consistent loss on intrusive metrics.

Load-bearing premise

The finding assumes that the 700-hour curated subset and the 2,500-hour full set differ mainly in how clean their labels are, but the filtering also changes which datasets, speakers, languages, and recording conditions are represented.

Editorial extensions

If this is right

  • For both discriminative and generative models, quality-ranked subsets outperform uniform random subsets of equal size on all reported non-intrusive metrics.
  • Training on the 700-hour filtered set improves perceptual quality scores over the full 2,500-hour set, while intrusive metrics stay on par or slightly worse.
  • Warm-starting from a full-data checkpoint and fine-tuning on the curated 700-hour subset yields the best results on multiple intrusive and non-intrusive metrics.
  • Scaling the amount of noise source from 50 to 200 hours shows no pronounced scaling curve on either the blind test set or an internal real-world test set, suggesting noise quantity alone is not the driver.

Reading between the lines

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

  • The paper's own composition table leaves open that part of the 700h-TBF advantage comes from dropping WSJ entirely and shrinking EARS from 107 to 16 hours, so a matched-composition control would separate quality from composition.
  • The filtering relies on the same non-intrusive metrics used for evaluation, so a fresh human-listening test would check whether the perceptual gains are artifacts of metric optimization.
  • If quality is the dominant scaling lever, the 60k-hour URGENT2025 Track2 corpus should be exploited by curation rather than by adding hours, a direction the paper names as future work.
  • The null noise-scaling result suggests future curation should target noise diversity, not just clean-label quality.
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. The paper argues that for scaling speech enhancement models, data quality matters more than data quantity. Using the URGENT2025 Track1 training set (2,500 hours of 'clean' speech), the authors score every utterance with five non-intrusive metrics (DNSMOS, NISQA, SIGMOS, UTMOS, Torchaudio-SQUIM-SDR) and apply dataset-specific thresholds to retain roughly 700 hours. They then train two model families (discriminative BSRNN and generative BSRNN-Flow) on the full set, the 700h filtered set, and on top-ranked 100h/350h subsets, and compare against uniformly random subsets of the same sizes. Table III reports that 700h-TBF beats 2500h-full on all six reported non-intrusive metrics for both model families, including DNSMOS Pro and DistillMOS that were not used in filtering, with intrusive metrics roughly on par. Figure 3 shows that top-ranked subsets generally outperform random subsets of equal size, and Section III-E reports flat noise-scaling curves. The paper concludes that curated smaller data can outperform much larger uncurated data, and proposes data curation as a key scaling strategy.

Significance. If the causal claim holds, the result is practically valuable for large-scale speech enhancement training: it would mean that inexpensive non-intrusive filtering can beat simply adding more data, and it would help explain the observed diminishing returns in the URGENT challenge. The study has real strengths: it tests two model families, reports consistent patterns across six non-intrusive metrics, and includes two held-out metrics not used in the filtering (DNSMOS Pro, DistillMOS), which partially addresses the circularity concern. The random-sampling baselines are a sensible design choice. However, the central causal interpretation ('quality, not quantity') is currently undermined by a composition confound, the absence of repeated-seed or error-bar analysis, and the overlap between filtering metrics and evaluation metrics.

major comments (3)
  1. [Section II-C and Table II] The causal claim that label quality, not dataset composition, drives the 700h-TBF advantage is not supported by the current experimental design. Table II shows that TBF changes the source mixture dramatically: WSJ is removed entirely, EARS drops from approximately 107h to 16h, Common Voice drops from approximately 1300h to 250h, and LibriVox drops from approximately 350h to 150h, while LibriTTS and VCTK retain larger fractions. The random-sampling baselines in Figure 3 control only for total hours, not for per-source composition, so the 700h-TBF vs 2500h-full comparison is confounded by speaker, language, recording-condition, and dataset-distribution differences. Please add a composition-matched control (for example, a stratified random subset of 700h drawn from the full set with the same source proportions as 700h-TBF, or per-source ablation experiments) and report how much of the Table III gain survives such a control.
  2. [Table I and Table III] Four of the six non-intrusive metrics reported in Table III (DNSMOS, NISQA, UTMOS, SIGMOS) are the same metrics whose thresholds define TBF in Table I, so part of the headline improvement is a selection effect on the evaluation instrument. The held-out DNSMOS Pro and DistillMOS columns are therefore crucial evidence, and the paper should quantify the gain on held-out versus filtering metrics explicitly. In addition, the thresholds in Table I are manually chosen per dataset, and the paper does not study how sensitive the 700h-TBF result is to these cutoffs. Please add a threshold-sensitivity analysis (for example, stricter and looser thresholds, or cross-validated threshold selection) and state whether the held-out metric improvements are of the same relative size as the filtering-metric improvements.
  3. [Table III and Figure 3] The paper reports no repeated seeds, no error bars, and no measure of variability. Each training condition appears to be a single run, and the random subsets are drawn once. Observed differences such as DNSMOS 2.80 vs 2.85, or SDR 10.89 vs 10.71, may be within run-to-run or subset-sampling variation, which is important given that the paper's claim is that 700h-TBF 'consistently outperform[s]' 2500h-full. Please provide at least two or three training seeds (or a nonparametric bootstrap over test utterances) for the headline 700h-TBF vs 2500h-full comparison and for the random baselines in Figure 3, and state whether the qualitative ordering holds under that variability.
minor comments (5)
  1. [Section III-E] The text contains a typo: 'URGRENT2025' should be 'URGENT2025'.
  2. [Table III caption] The caption spells 'DITILLMOS' but should be 'DistillMOS'.
  3. [Throughout] Dataset names are written with inconsistent spacing, e.g., 'V oiceBank', 'Common V oice', 'LibriV ox'; please use 'VoiceBank', 'Common Voice', and 'LibriVox' consistently.
  4. [Section II-C and Section III-D] The description of the top-ranked 700h subset as 'identical to 700h-TBF' is stated twice; please clarify whether this is true by construction or empirically because the top-700 ranking and the threshold filter happen to select the same utterances.
  5. [Section III-A] The high-pass filter at 75 Hz applied to the clean speech source before simulation is a preprocessing choice that could interact with the low-frequency noise issues described in Section II-B; a brief ablation or statement about whether it is applied uniformly to all training subsets would help.

Circularity Check

1 steps flagged · score 4.0 of 10

The '700h-TBF beats 2500h-full' result is partly a selection effect: the subset is defined by thresholds on DNSMOS/NISQA/UTMOS/SIGMOS, and Table III evaluates on those same four metrics; held-out DNSMOS Pro/DistillMOS give partial independent support, so the circularity is partial, not total.

  1. self definitional [Section II-C (Data Curation Strategy, Table I) and Section III-A / Table III (evaluation metrics)]
    ""We applied dataset-specific minimum quality thresholds using multiple non-intrusive metrics. Utterances failing to meet any threshold for their respective dataset were excluded. ... We report six non-intrusive metrics, including DNSMOS, NISQA, UTMOS, SIGMOS, DNSMOS Pro, and Distill-MOS in the evaluation, where the former four are used in the TBF, while the other two are not.""

    The 700h-TBF training subset is constructed by retaining exactly the utterances that pass the DNSMOS, SigMOS, UTMOS, NISQA (and SQUIM-SDR) thresholds of Table I. The paper's central evidence that this subset 'consistently outperform[s]' the 2500h-full set (Section III-C) is then taken from DNSMOS, NISQA, UTMOS, and SIGMOS—the same metrics used to define 'quality' in the curation step. Because the model is trained to reconstruct targets that were selected for high scores on these instruments, the four headline metrics are partly a re-measurement of the selection criterion rather than an independent test of it.

full rationale

The central derivation is not wholly circular: the paper reports two metrics not used in filtering (DNSMOS Pro, DistillMOS) that also improve, and intrusive results are on par or slightly worse, which is an honest non-intrusive-versus-intrusive tension. The self-citations to URGENT2025 [16] and prior scaling studies [13], [14] describe real external challenge/benchmark results rather than uniqueness theorems, so they are not load-bearing circularity by themselves. The composition shift between 2500h-full and 700h-TBF (Table II: WSJ dropped, EARS 107h to 16h, Common Voice 1300h to 250h) is a confound for the causal 'quality, not composition' claim, but it is a correctness threat, not a circularity. Score 4 reflects that part of the headline non-intrusive advantage reduces to the same instruments used to define the curated subset, while the held-out metrics and intrusive results keep the central claim from being fully definitional.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on six hand-chosen filtering thresholds, one hand-chosen high-pass cutoff, an assumed validity of non-intrusive metrics as quality ground truth, and an assumed causal role of label quality versus dataset composition. No new physical or model entities are introduced.

free parameters (6)
  • DNSMOS threshold = 2.5 (EARS), 3.0 (Common Voice ZH and others)
    Dataset-specific minimum in Table I; chosen by hand without optimization or theoretical basis, and directly controls which utterances remain in the curated subset.
  • SigMOS threshold = 2.5 (EARS), 3.0 (Common Voice ZH and others)
    Manually selected per dataset in Table I; used as one of five quality gates.
  • UTMOS threshold = 2.5 (EARS), 3.0 (Common Voice ZH and others)
    Manually selected per dataset in Table I; part of the filtering criterion.
  • NISQA threshold = 3.0 (EARS), 4.0 (Common Voice ZH and others)
    Manually selected per dataset in Table I; affects which utterances survive TBF.
  • SQUIM-SDR threshold = 0.0 (EARS and Common Voice ZH), 20 (others)
    Manually selected per dataset in Table I; the '20' for other datasets is a strong SI-SDR gate.
  • High-pass filter cutoff = 75 Hz
    Applied to the clean speech source before simulation in Section III-A to remove infrasound narrowband noise; chosen by hand and not ablated or optimized.
assumptions (4)
  • domain assumption Non-intrusive metrics (DNSMOS, NISQA, SIGMOS, SQUIM-SDR, UTMOS) are valid measures of clean-speech label quality and transfer to held-out evaluation.
    Used as the ground truth for curation in Section II-C and partly for evaluation in Section III-A; if these metrics misrank utterances, the 'high-quality' subset is not actually high quality.
  • domain assumption The dynamic simulation with the seven URGENT2025 distortions and the URGENT2025 blind test set represent general speech enhancement conditions.
    All scaling and quality conclusions are drawn on this protocol (Sections II-A and III-A); transfer to other noise and room impulse response distributions is untested.
  • ad hoc to paper Removing utterances below the thresholds improves performance because of label quality, not because of correlated changes in dataset composition.
    Table II shows large compositional changes (EARS from 107h to 16h, Common Voice from 1300h to 250h, WSJ dropped); the paper does not control for this, yet attributes the gains to quality.
  • domain assumption Single-talker clean speech is the correct training target, and multi-talker or distorted 'clean' labels are harmful defects.
    The paper treats multi-talker speech and various noises in clean labels as defects (Section II-B, Figure 1), but acknowledges the impact on generative versus discriminative training is uncertain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Less is More: Data Curation Matters in Scaling Speech Enhancement." pith.science (2026). https://pith.science/paper/OETW6E43

@misc{pith2026250623859,
  author       = {Pith},
  title        = {Pith review of: Less is More: Data Curation Matters in Scaling Speech Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OETW6E43}},
  note         = {Machine review of arXiv:2506.23859}
}
read the original abstract

The vast majority of modern speech enhancement systems rely on data-driven neural network models. Conventionally, larger datasets are presumed to yield superior model performance, an observation empirically validated across numerous tasks in other domains. However, recent studies reveal diminishing returns when scaling speech enhancement data. We focus on a critical factor: prevalent quality issues in ``clean'' training labels within large-scale datasets. This work re-examines this phenomenon and demonstrates that, within large-scale training sets, prioritizing high-quality training data is more important than merely expanding the data volume. Experimental findings suggest that models trained on a carefully curated subset of 700 hours can outperform models trained on the 2,500-hour full dataset. This outcome highlights the crucial role of data curation in scaling speech enhancement systems effectively.

Figures

Figures reproduced from arXiv: 2506.23859 by the authors.

Figure 1
Figure 1. Training target of low quality selected from Librivox [17], Common Voice [18], and VCKT [19] datasets, which are commonly used in speech [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Histogram of non-intrusive metrics of the target speech source on the full dataset and Threshold-Based-Filtered (TBF) dataset. It is noted that the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Impact of training data selection and scale on SE performance. Both discriminative BSRNN and generative BSRNN-Flow are evaluated. The X-axis [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Impact of noise source scale in the training data. The X-axis is the amount of training noise source. Intrusive and non-intrusive metrics are represented [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

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

    cs.SD 2025-06 conditional novelty 6.0 of 10

    A multi-metric learned quality model (Uni-VERSA-Ext) is used as a differentiable training loss for speech enhancement, with a regularization term to prevent adversarial exploitation.

Reference graph

Works this paper leans on

48 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    P. C. Loizou,Speech Enhancement: Theory and Practice. CRC Press, 2007

    work page 2007

  2. [2]

    A Regression Approach to Speech Enhancement Based on Deep Neural Networks,

    Y . Xu, J. Du, L.-R. Dai, and C.-H. Lee, “A Regression Approach to Speech Enhancement Based on Deep Neural Networks,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, pp. 7–19, Jan. 2015

    work page 2015

  3. [3]

    Supervised Speech Separation Based on Deep Learning: An Overview,

    D. Wang and J. Chen, “Supervised Speech Separation Based on Deep Learning: An Overview,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 26, pp. 1702–1726, Oct. 2018

    work page 2018

  4. [4]

    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 Transactions on Audio, Speech, and Language Processing, vol. 27, pp. 1256–1266, Aug. 2019

    work page 2019

  5. [5]

    DCCRN: Deep Complex Convolution Recurrent Network for Phase-Aware Speech Enhancement,

    Y . Hu, Y . Liu, S. Lv, M. Xing, S. Zhang, Y . Fu, J. Wu, B. Zhang, and L. Xie, “DCCRN: Deep Complex Convolution Recurrent Network for Phase-Aware Speech Enhancement,” inInterspeech 2020, pp. 2472– 2476, ISCA, 2020

    work page 2020

  6. [6]

    ICASSP 2023 Deep Noise Suppression Challenge,

    H. Dubey, A. Aazami, V . Gopal, B. Naderi, S. Braun, R. Cutler, A. Ju, M. Zohourian, M. Tang, M. Golestaneh, and R. Aichner, “ICASSP 2023 Deep Noise Suppression Challenge,”IEEE Open Journal of Signal Processing, vol. 5, pp. 725–737, 2024

    2023

  7. [7]

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

    W. Wang, W. Zhang, C. Li, J. Shi, S. Watanabe, and Y . Qian, “Improving speech enhancement with multi-metric supervision from learned quality assessment,”arXiv preprint arXiv:2506.12260, 2025

    work page Pith review arXiv 2025

  8. [8]

    Conditional Diffusion Probabilistic Model for Speech Enhancement,

    Y .-J. Lu, Z.-Q. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional Diffusion Probabilistic Model for Speech Enhancement,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 7402–7406, 2022

    work page 2022

Show all 48 references
  1. [9]

    Diffusion-Based Gener- ative Modeling With Discriminative Guidance for Streamable Speech Enhancement,

    C. Li, S. Cornell, S. Watanabe, and Y . Qian, “Diffusion-Based Gener- ative Modeling With Discriminative Guidance for Streamable Speech Enhancement,” in2024 IEEE Spoken Language Technology Workshop (SLT), pp. 333–340, Dec. 2024

    2024

  2. [10]

    Speech Enhancement and Dereverberation With Diffusion-Based Gen- erative Models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech Enhancement and Dereverberation With Diffusion-Based Gen- erative Models,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 2351–2364, 2023

    2023

  3. [11]

    FlowSE: Flow Matching- based Speech Enhancement,

    S. Lee, S. Cheong, S. Han, and J. W. Shin, “FlowSE: Flow Matching- based Speech Enhancement,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, Apr. 2025

    2025

  4. [12]

    Speech Enhancement for a Noise-Robust Text-to-Speech Synthesis System Using Deep Recurrent Neural Networks,

    C. Valentini-Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “Speech Enhancement for a Noise-Robust Text-to-Speech Synthesis System Using Deep Recurrent Neural Networks,” inProc. Interspeech 2016, pp. 352–356, 2016

    2016

  5. [13]

    Beyond Performance Plateaus: A Comprehensive Study on Scalability in Speech Enhancement,

    W. Zhang, K. Saijo, J.-w. Jung, C. Li, S. Watanabe, and Y . Qian, “Beyond Performance Plateaus: A Comprehensive Study on Scalability in Speech Enhancement,” inProc. Interspeech 2024, pp. 1740–1744, 2024

    2024

  6. [14]

    The Effect of Training Dataset Size on Discriminative and Diffusion-Based Speech Enhancement Systems,

    P. Gonzalez, Z.-H. Tan, J. Østergaard, J. Jensen, T. S. Alstrøm, and T. May, “The Effect of Training Dataset Size on Discriminative and Diffusion-Based Speech Enhancement Systems,”IEEE Signal Process- ing Letters, vol. 31, pp. 2225–2229, 2024

    2024

  7. [15]

    URGENT Challenge: Universality, Robustness, and Generalizability For Speech Enhancement,

    W. Zhang, R. Scheibler, K. Saijo, S. Cornell, C. Li, Z. Ni, A. Kumar, J. Pirklbauer, M. Sach, S. Watanabe, T. Fingscheidt, and Y . Qian, “URGENT Challenge: Universality, Robustness, and Generalizability For Speech Enhancement,” inInterspeech 2024, pp. 4868–4872, 2024

    2024

  8. [16]

    Interspeech 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, “Interspeech 2025 URGENT Speech Enhancement Challenge,” inInterspeech 2026, 2026

    2025

  9. [17]

    Librivox: Free public domain audiobooks,

    J. Kearns, “Librivox: Free public domain audiobooks,”Reference Re- views, vol. 28, no. 1, pp. 7–8, 2014

    2014

  10. [18]

    Common V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Hen- retty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common V oice: A Massively-Multilingual Speech Corpus,” inProceedings of the Twelfth Language Resources and Evaluation Conference(N. Calzolari, F. B ´echet, P....

    2020

  11. [19]

    The voice bank corpus: Design, collection and data analysis of a large regional accent speech database,

    C. Veaux, J. Yamagishi, and S. King, “The voice bank corpus: Design, collection and data analysis of a large regional accent speech database,” in2013 International Conference Oriental COCOSDA Held Jointly with 2013 Conference on Asian Spoken Language Research and Evaluation (O...

    2013

  12. [20]

    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

  13. [21]

    DNSMOS Pro: A Reduced-Size DNN for Probabilistic MOS of Speech,

    F. Cumlin, X. Liang, V . Ungureanu, C. K. A. Reddy, C. Sch ¨uldt, and S. Chatterjee, “DNSMOS Pro: A Reduced-Size DNN for Probabilistic MOS of Speech,” inProc. Interspeech 2024, pp. 4818–4822, 2024

    2024

  14. [22]

    Self-Supervised Speech Quality Estimation and Enhancement Using Only Clean Speech,

    S.-W. Fu, K.-H. Hung, Y . Tsao, and Y .-C. F. Wang, “Self-Supervised Speech Quality Estimation and Enhancement Using Only Clean Speech,” inThe Twelfth International Conference on Learning Representations, Oct. 2023

    2023

  15. [23]

    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 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, Apr. 2025

    2025

  16. [24]

    Dnsmos: A Non-Intrusive Per- ceptual Objective Speech Quality Metric to Evaluate Noise Suppressors,

    C. K. A. Reddy, V . Gopal, and R. Cutler, “Dnsmos: A Non-Intrusive Per- ceptual Objective Speech Quality Metric to Evaluate Noise Suppressors,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6493–6497, June 2021

    2021

  17. [25]

    EARS: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation,

    J. Richter, Y .-C. Wu, S. Krenn, S. Welker, B. Lay, S. Watanabe, A. Richard, and T. Gerkmann, “EARS: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation,” inProc. Interspeech 2024, pp. 4873–4877, 2024

    2024

  18. [26]

    BSS EV AL Toolbox User Guide – Revision 2.0,

    C. F ´evotte, R. Gribonval, and E. Vincent, “BSS EV AL Toolbox User Guide – Revision 2.0,” report, 2005

    2005

  19. [27]

    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

  20. [28]

    LibriTTS: A Corpus Derived from LibriSpeech for Text- to-Speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “LibriTTS: A Corpus Derived from LibriSpeech for Text- to-Speech,” inProc. Interspeech 2019, pp. 1526–1530, 2019

    2019

  21. [29]

    MLS: A Large-Scale Multilingual Dataset for Speech Research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “MLS: A Large-Scale Multilingual Dataset for Speech Research,” inProc. Interspeech 2020, pp. 2757–2761, 2020

    2020

  22. [30]

    CSR-I (WSJ0) Complete,

    Garofolo, John S., Graff, David, Paul, Doug, and Pallett, David, “CSR-I (WSJ0) Complete,” 2007

    2007

  23. [31]

    NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,

    G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,” inProc. Interspeech 2021, pp. 2127–2131, 2021

    2021

  24. [32]

    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

  25. [33]

    Torchaudio-Squim: Reference-Less Speech Quality and Intelligibility Measures in Torchaudio,

    A. Kumar, K. Tan, Z. Ni, P. Manocha, X. Zhang, E. Henderson, and B. Xu, “Torchaudio-Squim: Reference-Less Speech Quality and Intelligibility Measures in Torchaudio,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5...

    2023

  26. [34]

    SDR – Half- baked or Well Done?,

    J. L. Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR – Half- baked or Well Done?,” inICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 626–630, May 2019

    2019

  27. [35]

    An Algo- rithm for Intelligibility Prediction of Time–Frequency Weighted Noisy Speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An Algo- rithm for Intelligibility Prediction of Time–Frequency Weighted Noisy Speech,”IEEE Transactions on Audio, Speech, and Language Process- ing, vol. 19, pp. 2125–2136, Sept. 2011

    2011

  28. [36]

    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,” inProc. Interspeech 2022, pp. 4521–4525, 2022

    2022

  29. [37]

    TF-GridNet: Making time-frequency domain models great again for monaural speaker separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watan- abe, “TF-GridNet: Making time-frequency domain models great again for monaural speaker separation,” inICASSP 2023 - 2023 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023

    2023

  30. [38]

    Performance measurement in blind audio source separation,

    E. Vincent, R. Gribonval, and C. Fevotte, “Performance measurement in blind audio source separation,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, pp. 1462–1469, July 2006

    2006

  31. [39]

    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, pp. 380–391, Oct. 1976

    1976

  32. [40]

    Efficient Monaural Speech Enhancement with Universal Sample Rate Band-Split RNN,

    J. Yu and Y . Luo, “Efficient Monaural Speech Enhancement with Universal Sample Rate Band-Split RNN,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, June 2023

    2023

  33. [41]

    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, 2023

    1901

  34. [42]

    Complexity Scaling for Speech De- noising,

    H. Chen, J. Yu, and C. Weng, “Complexity Scaling for Speech De- noising,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 12276–12280, 2024

    2024

  35. [43]

    TFPSNet: Time-Frequency Domain Path Scanning Network for Speech Separation,

    L. Yang, W. Liu, and W. Wang, “TFPSNet: Time-Frequency Domain Path Scanning Network for Speech Separation,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6842–6846, 2022

    2022

  36. [44]

    ESPnet- SE: End-To-End Speech Enhancement and Separation Toolkit Designed for ASR Integration,

    C. Li, J. Shi, W. Zhang, A. S. Subramanian, X. Chang, N. Kamo, M. Hira, T. Hayashi, C. Boeddeker, Z. Chen, and S. Watanabe, “ESPnet- SE: End-To-End Speech Enhancement and Separation Toolkit Designed for ASR Integration,” in2021 IEEE Spoken Language Technology Workshop (SLT), p...

    2021

  37. [45]

    Towards Low-Distortion Multi-Channel Speech En- hancement: The ESPNET-Se Submission to the L3DAS22 Challenge,

    Y .-J. Lu, S. Cornell, X. Chang, W. Zhang, C. Li, Z. Ni, Z.-Q. Wang, and S. Watanabe, “Towards Low-Distortion Multi-Channel Speech En- hancement: The ESPNET-Se Submission to the L3DAS22 Challenge,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Sig...

    2022

  38. [46]

    Flow Matching for Generative Modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow Matching for Generative Modeling,” inThe Eleventh International Conference on Learning Representations, Sept. 2022

    2022

  39. [47]

    Score-Based Generative Modeling through Stochastic Differential Equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-Based Generative Modeling through Stochastic Differential Equations,” in9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, OpenReview.net, 2021

    2021

  40. [48]

    Scale this, not that: Investi- gating key dataset attributes for efficient speech enhancement scaling,

    L. Zhang, W. Zhang, C. Li, and Y . Qian, “Scale this, not that: Investi- gating key dataset attributes for efficient speech enhancement scaling,” arXiv preprint arXiv:2412.14890, 2024

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.