Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Interspeech 2025 URGENT Speech Enhancement Challenge

T0 review · 4 major / 6 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 Generative speech enhancement models can be language-dependent: in the URGENT Challenge, a purely generative system's character accuracy collapsed to 20% on Chinese and 37% on Japanese, while discriminative models stayed in a narrow band…

desk verdict Solid multilingual benchmark paper; the generative language-dependency claim is real but rests on one system and one metric, so it needs a careful rewrite. read the letter →

arxiv 2505.23212 v2 pith:HWMM6FVD submitted 2025-05-29 eess.AS

Kohei Saijo , Wangyou Zhang , Samuele Cornell , Robin Scheibler , Chenda Li , Zhaoheng Ni , Anurag Kumar , Marvin Sach
show 4 more authors
Yihui Fu Wei Wang Tim Fingscheidt Shinji Watanabe
This is my paper
classification eess.AS
keywords speechenhancementuniversallanguagedependencygenerativemultilingualbenchmarkqualityevaluationchallengeresultshallucination
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 describes the second URGENT speech enhancement challenge, built to test universal models that handle seven distortion types, multiple sampling rates, and multilingual speech. Its two headline findings are that a purely generative model can be strongly language-dependent — character accuracy fell to 20.1% on Chinese and 36.8% on Japanese while staying between 68.0% and 80.9% for European languages — and that some generative or hybrid systems were preferred by human listeners over the top discriminative system even though the discriminative system won the aggregate ranking. The paper concludes that evaluating generative speech enhancement requires content-aware metrics such as character accuracy, because non-intrusive quality scores keep looking high when the model hallucinates spoken content. The best overall system was purely discriminative, while most competitive submissions combined discriminative and generative losses.

What carries the argument

The instrument that carries the argument is the challenge protocol itself: a fixed distortion model applying up to five of seven distortions, multilingual corpora covering five training languages plus Japanese as an unseen test language, two training-data scales of approximately 2.5k and 60k hours, and a final ranking obtained by averaging ranks across 14 metrics in five categories. The language-dependency result rests on per-language comparison of two metrics in particular: character accuracy (1 − character error rate), computed by a multilingual automatic speech recognizer, and a non-intrusive speech-quality score. The generative model at the center of the finding is a latent-diffusion enhancer followed by vocoding; its behavior is contrasted with discriminative sub-band recurrent models and hybrid cascades.

What would settle it

Recompute the generative model's character accuracy on Japanese and Chinese using recognizers trained specifically on those languages, or with human transcripts, and compare with the multilingual recognizer's scores; if the gap to European languages largely disappears, the reported language dependency is an artifact of the recognizer rather than of the enhancement model. A complementary test is to train the same latent-diffusion architecture with Japanese data and see whether the dependency vanishes.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the choice of model family changes how a universal speech enhancer behaves across languages. In the challenge's blind test, the purely generative latent-diffusion system T13 produced character accuracy of 74.0% on German, 68.0% on English, 68.6% on French, and 80.9% on Spanish, but only 20.1% on Chinese and 36.8% on Japanese, the latter being a language unseen during training. Discriminative systems T1 and T10 stayed between 70.2% and 86.4% across all six languages. The paper also reports that T13 scored highest on subjective MOS (3.69 vs 3.24 for the top discriminative system) while ranking last on content preservation, and that its outputs sometimes hallucinated spoken content under low signal-to-noise ratios, with Japanese inputs occasionally sounding like English or other European languages. The challenge's aggregate ranking was nonetheless led by a pure discriminative model, with hybrids in the next positions.

Load-bearing premise

The language-dependency conclusion assumes that the same character-accuracy metric and the same non-intrusive quality metric are equally fair across all six languages; if those metrics are inherently harsher for Chinese or Japanese, the apparent collapse of the generative model would be partly a measurement artifact.

Editorial extensions

If this is right

  • Hybrid discriminative-plus-generative systems formed the bulk of the competitive field, but a purely discriminative model won the overall ranking.
  • Generative and hybrid systems can outscore the top discriminative model in subjective listening tests while losing on objective content metrics, so ranking depends on which metrics are used.
  • Purely generative speech enhancement should be treated as language-dependent until shown otherwise, and deployment on under-represented languages needs content-preservation checks.
  • Increasing training data from about 2.5k to 60k hours did not yield a system that beat the best small-track system, suggesting that raw data scale is not decisive when the added data contain noise.
  • Non-intrusive quality metrics alone cannot detect hallucinated speech content, so evaluating generative speech enhancement requires pairing them with recognition-based or content-similarity metrics.

Reading between the lines

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

  • The paper leaves open whether the language dependency is caused by the generative objective itself or by the English-heavy training distribution; a direct test would train the same latent-diffusion architecture on Japanese and measure whether the gap closes.
  • The subjective preference for generative models may partly be an artifact of the P.808 absolute category rating protocol, which asks about overall quality and does not direct listeners to check content fidelity; a content-aware listening protocol could reverse the MOS ordering.
  • If the language dependency replicates across generative architectures, it has a practical corollary for universal speech enhancement: content metrics like character accuracy should be reported per language in future benchmarks, not only as global averages.
  • The apparent failure of scaling from 2.5k to 60k hours suggests the limiting factor is data quality and filtering, not model capacity, which points to noisy-data utilization as a research direction the authors only begin to explore.
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 / 6 minor

Summary. This paper presents the Interspeech 2025 URGENT Challenge, the second edition of a benchmark for universal speech enhancement. It describes the task design (seven distortion types, multilingual training and blind-test data, and two tracks with roughly 2.5k and 60k hours of speech), a 14-metric evaluation protocol with a rank-based aggregation procedure, and results for 32 submissions. The main reported findings are that the best system is a purely discriminative model, that most competitive systems are hybrids, that some generative or hybrid systems receive higher English MOS than the top discriminative system, and that a purely generative system (T13) shows a large character-accuracy drop on Chinese and Japanese, which the paper interprets as evidence that purely generative SE models can exhibit language dependency. The paper also offers preliminary analyses of metric limitations and of the effect of adding noisy training data at scale.

Significance. If the findings hold, the challenge is a valuable community resource: it provides a reproducible evaluation framework, a public leaderboard, a strong baseline, a multilingual blind test that includes an unseen language, and a metric suite that explicitly tries to penalize content hallucination in generative enhancement. The authors are appropriately cautious in several places, noting that the subjective test is English-only, that data filtering is imperfect, and that the analysis is preliminary. The main scientific claim about language dependency is nevertheless supported by only one system and one downstream ASR metric, with no language-specific calibration, so the broader significance of that claim depends on additional validation. The data-scalability conclusion also needs a cross-track comparison on common metric values rather than within-track ranks.

major comments (4)
  1. [§4.2, Table 3] The headline finding that "purely generative SE models can exhibit language dependency" rests on a single system (T13) and a single downstream metric (CAcc with OWSM v3.1). The paper itself notes in §2 and §4.3 that objective metrics can be language dependent, but it does not provide a per-language calibration or a clean-speech ceiling for CAcc. The noisy baseline T22 shows cross-language CAcc variation (69.3–80.8) comparable to the discriminative models, so a uniform metric bias cannot by itself explain T13's collapse; however, a metric-model interaction (e.g., OWSM being particularly unforgiving of T13's vocoder artifacts on Chinese or Japanese) remains plausible. Please report per-language OWSM accuracy on clean references or corroborate CAcc with a second multilingual ASR metric and/or manual transcripts of T13's outputs for each language.
  2. [§4.2, Table 3] The prose states that T13's performance "markedly declines when applied to an unseen language, namely Japanese," but Table 3 shows T13's lowest CAcc is on Chinese (20.1%), and Chinese is among the languages available in the challenge training corpora (CommonVoice 19.0, Table 1). This discrepancy means the "unseen language" framing is not supported unless T13 was trained without Chinese, which the paper does not report. Please clarify T13's actual training language distribution and re-frame the claim accordingly.
  3. [§4.2, Table 3] The qualitative explanation that T13 "occasionally hallucinated spoken content, particularly under low-SNR conditions" is not quantified per language. Since this is the paper's main new scientific finding, please provide a per-language breakdown of hallucination frequency or content-preservation metrics (e.g., SBS or LPS by language) and the SNR distribution of the blind test, so that the Chinese/Japanese CAcc collapse can be attributed to content hallucination rather than to other artifacts.
  4. [§4.1, Table 2] The statement "no systems in Track2 outperform the best system in Track1" is not directly supported by Table 2 because the final ranking scores are average ranks computed within each track and are therefore not comparable across tracks; for example, Track2's T2' has a better ranking score (2.50) than Track1's T1 (2.97). The raw metric values mostly favor T1, but T2' has higher MOS and higher non-intrusive metric scores, so the cross-track comparison needs to be made on common metric values, and the data-scalability conclusion should be qualified accordingly.
minor comments (6)
  1. [Table 1 and §3.3.1] The language label "zn" appears for CommonVoice 19.0; this should likely be "zh" for Chinese.
  2. [§4.1] The sentence describing T3 says "a cascade of of D, G, and D systems"; the duplicated "of" should be removed.
  3. [Table 2] The notation "T*’" and the † marker for Track2 systems are confusing; please define the relationship between Track1 and Track2 system IDs more explicitly.
  4. [§3.5] The definition "CAcc, 1 - character error rate" should use a proper minus sign or the word "one minus" for clarity.
  5. [§4.3] The claim that "the other two non-intrusive metrics exhibited similar trends" is not substantiated in the paper; either include the analogous scores or state that they are available on the leaderboard.
  6. [§4.3] The sentence "Table 3 demonstrates that DNSMOS yields high scores even when hallucinations occur" is stronger than the evidence shown; Table 3 alone shows high DNSMOS with low CAcc, which is consistent with hallucination, but the wording should be adjusted unless per-sample hallucination annotations are provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an externally benchmarked challenge report whose claims rest on measured system outputs and independent metrics, not on self-referential derivation.

full rationale

This paper is a challenge report, not a derivation. Its central claims—that hybrid systems dominate competitive submissions, that a purely discriminative model won, that generative models were preferred in P.808 MOS, and that the generative system T13 showed language-dependent CAcc—are supported by externally measured outputs (Table 2 and Table 3) using published, independent evaluation tools (DNSMOS, NISQA, UTMOS, POLQA, PESQ, ESTOI, SDR, MCD, LSD, SBS with mHuBERT-147, LPS, SpkSim with RawNet3, CAcc with OWSM v3.1, and ITU-T P.808 crowdsourcing). The ranking procedure (Section 3.5) is transparent and pre-specified, and it is not fitted to the results. The language-dependency analysis is an empirical observation about T13's measured CAcc and DNSMOS scores; the paper does not derive this behavior from its own assumptions. Its own caveats—that metrics can be language dependent (Section 2, Section 4.3) and that subjective scores were collected only on English (footnote 2)—weaken the strength of the claim but do not make it circular. The self-citations ([6], [7]) describe prior URGENT challenge design and preliminary analyses that are external benchmarks or methodology descriptions, not load-bearing derivations of the present results. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. The claim that T13 hallucinated English-like content on Japanese inputs (Section 4.2) is presented as an observed, audio-demo-supported finding, and the paper even recommends caution about the metrics used for generative models. Therefore, no circular step can be exhibited, and the appropriate score is 0.

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

No free parameters are fitted in this benchmark report. The analysis relies on the stated assumptions about the validity of the aggregated ranking, the language-agnosticism of the metrics, and the adequacy of the subjective protocol. The paper explicitly flags the second and third assumptions as limitations.

assumptions (3)
  • domain assumption The equal-weight average of category-wise rankings across 14 metrics is a meaningful measure of overall system quality.
    Section 3.5 defines the final ranking as the average of category-wise rankings; no justification is given for equal weighting across metric categories, and the choice can change the ordering.
  • domain assumption Cross-language comparisons of CAcc and DNSMOS are valid without language-specific calibration.
    Section 4.2 compares DNSMOS and CAcc across six languages to infer model language dependency, while Sections 2 and 4.3 note that metrics themselves can be language-dependent and that non-intrusive metrics fail to penalize hallucination.
  • domain assumption The subjective P.808 MOS test is a suitable measure for generative SE quality in this comparison.
    The paper uses P.808 MOS to state that generative and hybrid approaches are preferred over the top discriminative model, while acknowledging the protocol does not penalize content hallucination (Section 4.3), making the preference partly an artifact of the metric.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interspeech 2025 URGENT Speech Enhancement Challenge." pith.science (2026). https://pith.science/paper/HWMM6FVD

@misc{pith2026250523212,
  author       = {Pith},
  title        = {Pith review of: Interspeech 2025 URGENT Speech Enhancement Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWMM6FVD}},
  note         = {Machine review of arXiv:2505.23212}
}
read the original abstract

There has been a growing effort to develop universal speech enhancement (SE) to handle inputs with various speech distortions and recording conditions. The URGENT Challenge series aims to foster such universal SE by embracing a broad range of distortion types, increasing data diversity, and incorporating extensive evaluation metrics. This work introduces the Interspeech 2025 URGENT Challenge, the second edition of the series, to explore several aspects that have received limited attention so far: language dependency, universality for more distortion types, data scalability, and the effectiveness of using noisy training data. We received 32 submissions, where the best system uses a discriminative model, while most other competitive ones are hybrid methods. Analysis reveals some key findings: (i) some generative or hybrid approaches are preferred in subjective evaluations over the top discriminative model, and (ii) purely generative SE models can exhibit language dependency.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. UniSE: A Unified Framework for Decoder-Only Autoregressive LM-Based Speech Enhancement

    cs.SD 2025-10 conditional novelty 6.0 of 10

    A 63M-parameter decoder-only LM, UniSE, unifies speech restoration, target speaker extraction, and speech separation by generating BiCodec discrete tokens under task-specific prompts.

  2. AVFSNet: Audio-Visual Speech Separation for Flexible Number of Speakers with Multi-Scale and Multi-Task Learning

    eess.AS 2025-07 conditional novelty 6.0 of 10

    AVFSNet combines a Branchformer audio encoder, parallel audio-visual separation branches, and a mask-based counter to separate flexible numbers of speakers from audio-visual mixtures.

Reference graph

Works this paper leans on

68 extracted references · 60 canonical work pages · cited by 2 Pith papers

  1. [1]

    It has seen significant progress thanks to neural networks

    Introduction Speech enhancement (SE) aims to improve speech quality de- graded by noise, reverberation, or other distortions. It has seen significant progress thanks to neural networks. However, most prior work has focused on matched training–inference conditions and limited tasks such as noise suppression and dereverberation. In recent years, the advent ...

    work page 2024

  2. [2]

    The URGENT challenge complements existing ones by focusing on universality, generalizability, and robustness across diverse scenarios and evaluation metrics

    Related works Existing SE challenges have greatly advanced SE studies for specific scenarios, such as denoising and dereverberation [9, 10], speech restoration [11, 12], packet loss concealment [1], and so on. The URGENT challenge complements existing ones by focusing on universality, generalizability, and robustness across diverse scenarios and evaluatio...

    arXiv 2025

  3. [3]

    Task definition In the URGENT challenge, the SE process is defined as [6]: ˆx= SE(F(x)),(1) where x and ˆx are the desired and enhanced speech signals

    Challenge Design 3.1. Task definition In the URGENT challenge, the SE process is defined as [6]: ˆx= SE(F(x)),(1) where x and ˆx are the desired and enhanced speech signals. F(·) is the distortion model that degrades the desired signal. Our definition differs from the commonly adopted one in the literature in that the SE model has to handle (i) inputs wit...

    work page 2018

  4. [4]

    Non-intrusive SE metrics: DNSMOS [8], NISQA [16], and UTMOS [17]

    work page

  5. [5]

    Intrusive SE metrics: perceptual objective listening quality assessment (POLQA) [39], perceptual evaluation of speech quality (PESQ) [40], extended short-time objective intelli- gibility (ESTOI) [41], signal-to-distortion ratio (SDR) [42], mel cepstral distortion (MCD) [43], and log-spectral distance (LSD) [44]

    work page

  6. [6]

    Downstream-task-independent metrics: Levenshtein phone similarity (LPS) [45] and SpeechBERTScore (SBS) [46] with mHuBERT-147 [47]

    work page

  7. [7]

    Downstream-task-dependent metrics: speaker similarity (SpkSim) with RawNet3 [48] and character accuracy (CAcc, 1 - character error rate) with OWSM v3.1 [49]

    work page

  8. [8]

    Subjective metric: absolute category rating mean opinion score (MOS) via ITU-T P.808 test [50, 51], implemented on the Amazon Mechanical Turk (on only English data2). Compared to the first challenge [6], we made the following changes to support multilingual data: (i) we use an SSL model trained on multilingual data, mHuBERT-147, in SBS, and (ii) we evalua...

    work page

Show all 68 references
  1. [9]

    Calculate the ranking for each metric

  2. [10]

    Average the rankings of the metrics in each category

  3. [11]

    hallucination

    Average the category-wise rankings obtained in Step 2. The overall ranking scores obtained in Step 3 are used to deter- mine the final ranking. 7Here, "hallucination" refers to discrepancies in spoken content or speaker characteristics between the noisy speech and enhanced spe...

  4. [12]

    How do you rate the overall quality of the following speech sample?

    Results In this section, we report the evaluation results of the submitted systems. We also provide some preliminary analysis based on the brief system descriptions we collected from participants. 4.1. Overall results Table 2 shows the evaluation results, where 12 out of 22 sy...

  5. [13]

    The most popular approach was hybrid models but purely discriminative models gave the best score

    Conclusion We introduced the Interspeech 2025 URGENT Challenge to in- vestigate language dependency, universality on various distortion types, the effectiveness of using noisy data, and the data scalabil- ity of SE models. The most popular approach was hybrid models but purely...

    2025

  6. [14]

    Acknowledgement This work was partially supported by JSPS KAKENHI Grant Number JP24KJ2096. The leaderboard evaluation has been supported by the PSC Bridges2 system via ACCESS allocation CIS210014, supported by National Science Foundation grants #2138259, #2138286, #2138307, #2...

  7. [15]

    INTERSPEECH 2022 audio deep packet loss concealment challenge,

    L. Dieneret al., “INTERSPEECH 2022 audio deep packet loss concealment challenge,” inProc. Interspeech, 2022, pp. 580–584

    2022

  8. [16]

    Wind noise reduction with a diffusion- based stochastic regeneration model,

    J.-M. Lemercieret al., “Wind noise reduction with a diffusion- based stochastic regeneration model,” inSpeech Communication; 15th ITG Conference. VDE, 2023, pp. 116–120

    2023

  9. [17]

    Universal speech enhancement with score-based diffusion,

    J. Serràet al., “Universal speech enhancement with score-based diffusion,”arXiv preprint arXiv:2206.03065, 2022

    2022 arXiv

  10. [18]

    Universal score-based speech enhancement with high content preservation,

    R. Scheibleret al., “Universal score-based speech enhancement with high content preservation,” inProc. Interspeech, 2024, pp. 1165–1169

    2024

  11. [19]

    Toward universal speech enhancement for diverse input conditions,

    W. Zhanget al., “Toward universal speech enhancement for diverse input conditions,” inProc. IEEE ASRU, 2023

    2023

  12. [20]

    URGENT challenge: Universality, robustness, and generalizability for speech enhancement,

    W. Zhanget al., “URGENT challenge: Universality, robustness, and generalizability for speech enhancement,” inProc. Interspeech, 2024, pp. 4868–4872

    2024

  13. [21]

    Lessons learned from the URGENT 2024 speech enhancement challenge,

    W. Zhanget al., “Lessons learned from the URGENT 2024 speech enhancement challenge,” inAccepted by Interspeech, 2025

    2024

  14. [22]

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

    C. K. Reddyet al., “DNSMOS P.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in Proc. IEEE ICASSP, 2022, pp. 886–890

    2022

  15. [23]

    The INTERSPEECH 2020 deep noise sup- pression challenge: Datasets, subjective testing framework, and challenge results,

    C. K. Reddyet al., “The INTERSPEECH 2020 deep noise sup- pression challenge: Datasets, subjective testing framework, and challenge results,” inProc. Interspeech, 2020, pp. 2492–2496

    2020

  16. [24]

    ICASSP 2023 deep noise suppression challenge,

    H. Dubeyet al., “ICASSP 2023 deep noise suppression challenge,” IEEE Open Journal of Signal Processing, pp. 1–13, 2024

    2023

  17. [25]

    ICASSP 2023 speech signal improvement chal- lenge,

    R. Cutleret al., “ICASSP 2023 speech signal improvement chal- lenge,”IEEE Open Journal of Signal Processing, pp. 1–12, 2024

    2023

  18. [26]

    ICASSP 2024 speech signal improvement challenge,

    N.-C. Risteaet al., “ICASSP 2024 speech signal improvement challenge,”IEEE Open Journal of Signal Processing, vol. 6, pp. 238–246, 2025. [Online]. Available: https://ieeexplore.ieee.org/ document/10830509

    2024

  19. [27]

    Disentangling the impacts of language and channel variability on speech separation networks,

    F.-L. Wanget al., “Disentangling the impacts of language and channel variability on speech separation networks,” inProc. Inter- speech, 2022, pp. 5343–5347

    2022

  20. [28]

    A methodology for improving PESQ accuracy for Chinese speech,

    F. L. Chonget al., “A methodology for improving PESQ accuracy for Chinese speech,” inTENCON 2005-2005 IEEE Region 10 Conference. IEEE, 2005, pp. 1–6

    2005

  21. [29]

    D. Konaneet al., “Impact of languages and accent on perceived speech quality predicted by perceptual evaluation of speech qual- ity (PESQ) and perceptual objective listening quality assessment (POLQA): Case of Moore, Dioula, French and English,”Open Journal of Applied Sciences...

    2021

  22. [30]

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

    G. Mittaget al., “NISQA: A deep CNN-self-attention model for multidimensional speech quality prediction with crowdsourced datasets,” inProc. Interspeech, 2021, pp. 2127–2131

    2021

  23. [31]

    UTMOS: UTokyo-SaruLab system for V oiceMOS challenge 2022,

    T. Saekiet al., “UTMOS: UTokyo-SaruLab system for V oiceMOS challenge 2022,” inProc. Interspeech, 2022, pp. 4521–4525

    2022

  24. [32]

    Noisy-target training: A training strategy for dnn-based speech enhancement without clean speech,

    T. Fujimuraet al., “Noisy-target training: A training strategy for dnn-based speech enhancement without clean speech,” inProc. EUSIPCO. IEEE, 2021, pp. 436–440

    2021

  25. [33]

    Unsupervised sound separation using mixture invariant training,

    S. Wisdomet al., “Unsupervised sound separation using mixture invariant training,”Proc. NIPS, vol. 33, pp. 3846–3857, 2020

    2020

  26. [34]

    Scaling laws for neural language models,

    J. Kaplanet al., “Scaling laws for neural language models,”arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  27. [35]

    Scaling vision transformers,

    X. Zhaiet al., “Scaling vision transformers,” inProc. CVPR, 2022, pp. 12 104–12 113

    2022

  28. [36]

    Beyond performance plateaus: A comprehensive study on scalability in speech enhancement,

    W. Zhanget al., “Beyond performance plateaus: A comprehensive study on scalability in speech enhancement,” inProc. Interspeech, 2024, pp. 1740–1744

    2024

  29. [37]

    LibriTTS: A corpus derived from LibriSpeech for text-to-speech,

    H. Zenet al., “LibriTTS: A corpus derived from LibriSpeech for text-to-speech,” inProc. Interspeech, 2019, pp. 1526–1530

    2019

  30. [38]

    The V oice Bank corpus: Design, collection and data analysis of a large regional accent speech database,

    C. Veauxet al., “The V oice Bank corpus: Design, collection and data analysis of a large regional accent speech database,” inProc. O-COCOSDA/CASLRE, 2013, pp. 1–4

    2013

  31. [39]

    LDC,LDC Catalog: CSR-I (WSJ0) Complete, University of Penn- sylvania, 1993

    1993

  32. [40]

    Philadelphia: Linguistic Data Consortium,LDC Catalog: CSR-II (WSJ1) Complete LDC94S13A, 1994

    1994

  33. [41]

    EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,

    J. Richteret al., “EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,” in Proc. Interspeech, 2024, pp. 4873–4877

    2024

  34. [42]

    MLS: A large-scale multilingual dataset for speech research,

    V . Pratapet al., “MLS: A large-scale multilingual dataset for speech research,” inProc. Interspeech, 2020, pp. 2757–2761

    2020

  35. [43]

    Common voice: A massively-multilingual speech corpus,

    R. Ardilaet al., “Common voice: A massively-multilingual speech corpus,” inProc. LREC, 2020, pp. 4218–4222

    2020

  36. [44]

    WHAM!: Extending speech separation to noisy environments,

    G. Wichernet al., “WHAM!: Extending speech separation to noisy environments,” inProc. Interspeech, 2019, pp. 1368–1372

    2019

  37. [45]

    FSD50K: an open dataset of human-labeled sound events,

    E. Fonsecaet al., “FSD50K: an open dataset of human-labeled sound events,”IEEE/ACM Trans. ASLP ., vol. 30, pp. 829–852, 2021

    2021

  38. [46]

    FMA: A dataset for music analysis,

    M. Defferrardet al., “FMA: A dataset for music analysis,”arXiv preprint arXiv:1612.01840, 2016

    2016 arXiv

  39. [47]

    The diverse environments multi-channel acous- tic noise database (DEMAND): A database of multichannel envi- ronmental noise recordings,

    J. Thiemannet al., “The diverse environments multi-channel acous- tic noise database (DEMAND): A database of multichannel envi- ronmental noise recordings,” inProc. Mtgs. Acoust., vol. 19, no. 1. AIP Publishing, 2013

    2013

  40. [48]

    A multi-device dataset for urban acoustic scene classification,

    A. Mesaroset al., “A multi-device dataset for urban acoustic scene classification,” inDCASE, 2018, pp. 9–13

    2018

  41. [49]

    Brudex database: Binaural room impulse re- sponses with uniformly distributed external microphones,

    D. Fejginet al., “Brudex database: Binaural room impulse re- sponses with uniformly distributed external microphones,” in Speech Communication; 15th ITG Conf., 2023, pp. 126–130

    2023

  42. [50]

    MYRiAD: A multi-array room acoustic database,

    T. Dietzenet al., “MYRiAD: A multi-array room acoustic database,”EURASIP J. Audio Speech Music Process., vol. 2023, article no. 17, pp. 1–14, Apr. 2023

    2023

  43. [51]

    High fidelity neural audio compression,

    A. Défossezet al., “High fidelity neural audio compression,”Trans- actions on Machine Learning Research, 2023

    2023

  44. [52]

    High-fidelity audio compression with improved rvqgan,

    R. Kumaret al., “High-fidelity audio compression with improved rvqgan,”Advances in Neural Information Processing Systems, vol. 36, pp. 27 980–27 993, 2023

    2023

  45. [53]

    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. Beerendset al., “Perceptual objective listening quality assess- ment (POLQA), the third generation ITU-T standard for end-to-end speech quality measurement part I–—temporal alignment,”Journal of The Audio Eng. Society, vol. 61, no. 6, pp. 366–384, 2013

    2013

  46. [54]

    Perceptual evaluation of speech quality (PESQ)— a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rixet al., “Perceptual evaluation of speech quality (PESQ)— a new method for speech quality assessment of telephone networks and codecs,” inProc. IEEE ICASSP, vol. 2, 2001, pp. 749–752

    2001

  47. [55]

    An algorithm for predicting the intelligi- bility of speech masked by modulated noise maskers,

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

    2009

  48. [56]

    Performance measurement in blind audio source separation,

    E. Vincentet al., “Performance measurement in blind audio source separation,”IEEE Trans. ASLP ., vol. 14, no. 4, pp. 1462–1469, 2006

    2006

  49. [57]

    Mel-cepstral distance measure for objective speech quality assessment,

    R. Kubichek, “Mel-cepstral distance measure for objective speech quality assessment,” inProc. of IEEE Pacific Rim Conf. on Com- munications Computers and Signal Process., 1993, pp. 125–128

    1993

  50. [58]

    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

  51. [59]

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

    J. Pirklbaueret al., “Evaluation metrics for generative speech enhancement methods: Issues and perspectives,” inSpeech Com- munication; 15th ITG Conference, 2023, pp. 265–269

    2023

  52. [60]

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

    T. Saekiet al., “SpeechBERTScore: Reference-aware automatic evaluation of speech generation leveraging NLP evaluation metrics,” inProc. Interspeech, 2024, pp. 4943–4947

    2024

  53. [61]

    mHuBERT-147: A compact multilingual HuBERT model,

    M. Zanon Boitoet al., “mHuBERT-147: A compact multilingual HuBERT model,” inProc. Interspeech, 2024, pp. 3939–3943

    2024

  54. [62]

    Pushing the limits of raw waveform speaker recognition,

    J.-w. Junget al., “Pushing the limits of raw waveform speaker recognition,” inProc. Interspeech, 2022, pp. 2228–2232

    2022

  55. [63]

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

    Y . Penget al., “OWSM v3.1: Better and faster open Whisper-style speech models based on e-branchformer,” inProc. Interspeech, 2024, pp. 352–356

    2024

  56. [64]

    ITU-T recommendation P.808, subjective evaluation of speech quality with a crowdsourcing approach,

    “ITU-T recommendation P.808, subjective evaluation of speech quality with a crowdsourcing approach,” ITU-T, Jun. 2018

    2018

  57. [65]

    An open source implementation of ITU- T recommendation P.808 with validation,

    B. Naderi and R. Cutler, “An open source implementation of ITU- T recommendation P.808 with validation,” inProc. Interspeech, 2020, pp. 2862–2866

    2020

  58. [66]

    The use of ranks to avoid the assumption of normal- ity implicit in the analysis of variance,

    M. Friedman, “The use of ranks to avoid the assumption of normal- ity implicit in the analysis of variance,”Journal of the American Statistical Association, vol. 32, no. 200, pp. 675–701, 1937

    1937

  59. [67]

    TF-GridNet: Integrating full-and sub-band modeling for speech separation,

    Z.-Q. Wanget al., “TF-GridNet: Integrating full-and sub-band modeling for speech separation,”IEEE/ACM Trans. ASLP ., vol. 31, pp. 3221–3236, 2023

    2023

  60. [68]

    ESPnet-SE: End-to-end speech enhancement and separation toolkit designed for ASR integration,

    C. Liet al., “ESPnet-SE: End-to-end speech enhancement and separation toolkit designed for ASR integration,” inProc. IEEE SLT, 2021, pp. 785–792

    2021

Pith tools

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