Pith. sign in

REVIEW 5 major objections 5 minor 36 references

TS-URGENet: A Three-stage Universal Robust and Generalizable Speech Enhancement Network

T0 review · 5 major / 5 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 three-stage pipeline for universal speech enhancement ranked 2nd in the 2025 URGENT Challenge.

desk verdict Credible 2nd-place challenge system, but the three-stage ordering is never ablated and the metric-aware losses partly optimize the reported metrics. read the letter →

arxiv 2505.18533 v1 pith:T3QDSTFQ submitted 2025-05-24 eess.AS cs.AI

Xiaobin Rong , Dahan Wang , Qinwen Hu , Yushi Wang , Yuxiang Hu , Jing Lu This is my paper · ORCID
classification eess.AScs.AI
keywords universalspeechenhancementrestorationthree-stagearchitecturepacketlossconcealmentbandwidthextensionmetric-awarefine-tuningGAN-basedTF-GridNet
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 reports a universal speech enhancement system that handles seven distortion types and seven input sampling rates with one model. The author's central claim is that the job should be split into three ordered stages—filling packet-loss gaps, separating speech from noise/reverberation/clipping/wind, and restoring bandwidth/codec/residual packet-loss quality—rather than attempting all distortions in a single pass. On top of this pipeline, a metric-aware fine-tuning loss directly optimizes the challenge's evaluation metrics, and a joint stage-and-metric fine-tuning step balances intrusive and subjective measures. The system placed 2nd in Track 1 of the Interspeech 2025 URGENT Challenge, which the authors offer as evidence that the decomposition works.

What carries the argument

The load-bearing mechanism is the ordered three-stage pipeline—filling (TF-GridNet-S with a residual connection, trained with a GAN loss), separation (TF-GridNet-L), and restoration (CWS-TF-GridNet, where channel-wise subband splitting and merging operate at a fixed 48 kHz with resampling at the stage boundaries). Around this pipeline sit two training devices: metric-aware fine-tuning (MAFT), which appends differentiable surrogate losses for MCD, PESQ, UTMOS, DNSMOS, and WavLM representation similarity, and joint stage-and-metric-aware fine-tuning (JFT), which replaces plain stage-aware fine-tuning with an adversarial and feature-matching loss added to the metric-aware objective. These mechanisms together let one 30.4 M parameter model adapt to 8, 16, 22.05, 24, 32, 44.1, and 48 kHz input while addressing all seven distortions.

What would settle it

Train the same three-stage model with the filling and separation stages swapped (or with the seven distortions regrouped across stages) while keeping total parameters, losses, and training data fixed, and compare on the official 1,000-pair validation set; if the reordered model matches or beats the reported metrics, the specific ordering is not the cause of the improvement.

Watch

Extended reading notes

Core claim

The central discovery is that a three-stage filling-separation-restoration architecture, built from TF-GridNet blocks with a channel-wise subband module in the final stage, can serve as a universal speech enhancer across diverse distortions and sample rates. The filling stage runs first because packet-loss discontinuities degrade the separation module; the separation stage suppresses noise, reverberation, clipping, and wind; and the restoration stage, operating at 48 kHz after resampling, handles bandwidth limitation, codec artifacts, and residual packet loss. Metric-aware fine-tuning adds differentiable terms for MCD, PESQ, UTMOS, DNSMOS, and WavLM-based representation similarity to the conventional reconstruction loss, and joint stage-plus-metric fine-tuning (JFT) keeps non-intrusive metrics high without collapsing intrusive ones. The authors find that each added stage helps on the official validation set, and the full system improves on the challenge baseline across nearly all reported metrics on the blind test set.

Load-bearing premise

The load-bearing premise is that grouping the seven distortions into a fixed order—packet loss first, then noise/reverberation/clipping/wind, then bandwidth/codec/residual packet loss—is what produces the gains, rather than the extra parameters, GAN training, or metric-aware losses that come with the larger three-stage system.

Editorial extensions

If this is right

  • A single universal enhancer can outperform dedicated per-distortion systems by ordering operations as fill-then-separate-then-restore instead of mixing all distortions in one network.
  • Filling packet loss before denoising improves NISQA, UTMOS, PESQ, MCD, and LSD over the same pipeline without the filling stage, so continuity of the input matters for downstream separation.
  • Metric-aware fine-tuning can push non-intrusive metrics up, but added loss terms conflict: the gains from the UTMOS-aware term came partly at the cost of SDR, MCD, LSD, and CAcc.
  • Joint stage-and-metric-aware fine-tuning (JFT) recovers much of the PESQ lost by plain stage-aware fine-tuning while keeping the NISQA and UTMOS gains.
  • Choosing the restoration module per input type, called bandwidth-aware inference, lets one system trade intrusive metrics for subjective quality; the authors selected this variant as their final submission.

Reading between the lines

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

  • A reader wanting to test the ordering hypothesis could swap the filling and separation stages or regroup the seven distortions while holding parameter count and losses fixed; the paper reports only whole-configuration ablations, so the specific stage assignment is a design hypothesis rather than a proven cause.
  • If the ordering is what carries the gain, the same fill-separate-restore logic may transfer to lighter backbones or streaming architectures, since the choice of TF-GridNet is not essential to the decomposition itself.
  • The metric-aware fine-tuning recipe is plug-and-play: any downstream metric with a differentiable proxy or pretrained representation model could be added to $L_2$ in place of UTMOS or WavLM, suggesting a general way to co-optimize a whole measurement suite.
  • Bandwidth-aware inference suggests a practical pattern for handling format heterogeneity: one model version per input condition, selected at inference, rather than a single unconditioned network.
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

5 major / 5 minor

Summary. The paper presents TS-URGENet, a submission to the Interspeech 2025 URGENT Challenge Track 1. The system uses a three-stage cascade: a filling stage for packet loss, a separation stage for noise, reverberation, clipping, and wind noise, and a restoration stage for bandwidth limitation, codec artifacts, and residual packet loss. Each stage is built on TF-GridNet variants, with GAN-based training and a metric-aware fine-tuning (MAFT) stage followed by joint stage-and-metric-aware fine-tuning (JFT). The authors report a 2nd-place ranking in Track 1 and provide ablations on the official validation set plus results on the official blind test set.

Significance. If the central claims are supported, the paper would demonstrate that an ordered three-stage decomposition of universal speech enhancement, combined with direct optimization of target metrics, can achieve top-tier challenge performance. The paper is valuable as a detailed system description: it gives stage-wise training configurations, loss weights, hyperparameters, and systematic ablation rows on the official validation set, which is a useful level of reproducibility for a challenge-system paper. The externally provided blind-test ranking is a credible strong outcome. However, the load-bearing attribution of the gains to the specific filling–separation–restoration ordering is not established by the reported ablations, and the metric-aware losses directly fit several of the metrics used for evaluation, so the mechanistic novelty and the independence of the quality gains both require substantial qualification.

major comments (5)
  1. [§2, Table 3] The paper's central novelty is the specific filling→separation→restoration ordering and the assignment of the seven distortion types to these stages, but Table 3 never varies this commitment. The ablation only adds or removes whole stages (ID1 vs ID2, ID6 vs ID9) or changes the loss and fine-tuning scheme (ID3–ID12); there is no permutation such as separation→filling→restoration, no regrouping of distortions across modules, and no single-stage baseline with matched parameter count or compute. The observed gains could therefore be caused by added parameters, additional GAN discriminators, or the higher total computation rather than by the proposed decomposition. This is load-bearing for the main contribution; the authors should add at least one order- or grouping-ablation, or substantially soften the attribution claim.
  2. [§2.2, Eqs. (8)–(9), Table 4] The metric-aware loss explicitly includes differentiable PESQ, UTMOS, DNSMOS, and WavLM-based terms, and Table 4 then reports improvements on PESQ, UTMOS, DNSMOS, and WavLM-derived metrics such as PhnSim and SpkSim. These gains are partly constructed by fitting the same evaluation metrics, so they cannot be treated as independent evidence of quality. The paper should state this circularity explicitly and identify which reported improvements are on metrics not used in training (e.g., SDR, ESTOI, LSD, CAcc), so the reader can separate constructed gains from generalization gains.
  3. [§2.1] The filling-stage section concedes that 'the model often fills the gaps with noise rather than valid speech,' which directly contradicts the abstract's claim that the filling stage 'mitigates packet loss by preliminarily filling lost regions' and the design goal of restoring the lost components. The fallback explanation—that continuity helps the subsequent separation stage—is plausible but is asserted rather than tested. A controlled comparison replacing the learned filling stage with a simple placeholder or interpolation, while keeping the same downstream stages, is needed to validate this design decision.
  4. [Table 3] Many of the headline differences that support the architectural claims are small, for example DNSMOS 3.05→3.08, UTMOS 2.47→2.54, and SDR 13.03→13.19 dB between ID1 and ID2. No confidence intervals, significance tests, or repeated-run variability are reported. Because these small differences carry the central attribution claims, the authors should report bootstrap intervals or statistical significance on the official validation set.
  5. [§4.1.3, IDs 10–12] The choice of JFT over SAFT and the selection of bandwidth-aware inference as the final submission are partly justified by the statement that SAFT 'delivers a clear improvement in subjective listening quality, which cannot be fully captured in the table.' No listening test, MOS panel, or any other subjective evidence is reported in the paper. Either report the subjective evaluation that supports this selection, or explicitly state that the decision rests on informal listening and is therefore not an empirical result of this paper.
minor comments (5)
  1. [Table 3] There are typos in the table: 'Filll+Sep-MAFTv5' should be 'Fill+Sep-MAFTv5', and IDs 11–12 use 'MATFv3' where the text and other rows use 'MAFTv3'.
  2. [References] References [3] and [4] are identical; one should be replaced with the appropriate distinct citation or removed.
  3. [§2.2] Equation (7) uses log10 of a sigmoid of cosine similarity; the notation is unusual and the domain of the argument is not discussed. A brief explanation of this design choice would improve readability.
  4. [§4.1.3] The bandwidth-aware inference rule selects between two restoration modules based on whether speech is identified as bandwidth-limited, but the detection method is not described. Please specify how this identification is performed.
  5. [§2] The statement that TF-GridNet 'inherently satisfies the sampling-frequency-independent requirement' would benefit from a citation or a one-sentence justification, since this property is used to justify the flexibility of the filling and separation stages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the metric-aware losses are transparent training objectives, not independent predictions, and the paper's headline result is an external challenge ranking rather than a circular derivation.

full rationale

The paper makes no first-principles prediction whose output is equivalent to an input by construction. Its headline claim is the external URGENT 2025 Challenge result, which is an independent, externally adjudicated outcome. The metric-aware losses in Eqs. 8-9 explicitly include differentiable PESQ, UTMOS, DNSMOS, and WavLM terms, and Tables 3-4 report those same metrics; this is an intentional optimization objective, transparently labeled as 'metric-aware fine-tuning', not a hidden or renamed input. Reporting gains on a training objective is an empirical engineering result, not a circular derivation, and the final blind-test evaluation is a genuinely held-out set. The self-citation [16] appears only in a related-work list of two-stage approaches and is not load-bearing. The lack of ablations that vary stage order or distortion grouping is a limitation on causal attribution of the architecture's benefit, but it is not circularity. Accordingly, no specific circular step can be exhibited under the required standard.

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

The central claims rest on hand-chosen loss weights, an assumed stage-wise distortion grouping, and a benchmark-specific evaluation. No new physical entities are introduced; CWS-TF-GridNet is a composition of existing modules.

free parameters (4)
  • L1 loss weights (LSDR, LLSD, Lmag, Lreal, Limag weights) = 2, 1.5, 70, 30, 30
    Hand-chosen coefficients in Eq. 1; no sensitivity analysis is reported.
  • L2 metric-aware loss weights (MCD, PESQ, UTMOS, DNSMOS, WavLM) = 0.004, 0.5, 0.5, 0.4, 2.5
    Hand-tuned weights in Eqs. 8-9; the paper says an exhaustive investigation was done but reports only the chosen values.
  • L3 joint fine-tuning weights = 10, 1, 0.2
    Coefficients in Eq. 10 for the JFT loss combining L2, adversarial, and feature-matching terms.
  • GAN loss weights (reconstruction, adversarial, feature matching) = 20, 1, 1
    Weights used in filling and restoration stages, inherited from Gesper [13] without re-tuning.
assumptions (4)
  • ad hoc to paper The assignment of each distortion type to a specific stage (filling, separation, restoration) is a valid decomposition that does not harm cross-distortion interactions.
    Sec. 2 partitions the seven distortions by assumed characteristics; no search over alternative groupings or stage order is reported.
  • domain assumption Differentiable versions of PESQ, DNSMOS, UTMOS, and WavLM-based metrics are faithful surrogates whose optimization improves the true scores.
    Sec. 2.2 uses these as loss terms; the paper does not validate that optimizing the surrogates transfers beyond the challenge's own scoring scripts.
  • domain assumption Filling packet-loss gaps with noise (rather than speech) improves downstream separation by maintaining signal continuity.
    Sec. 2.1 admits the model fills gaps with noise but claims this helps continuity; the supportive evidence is the ID1-vs-ID2 ablation, which changes both the stage and the model.
  • domain assumption The official URGENT 2025 simulation scripts and the blind test set are representative of real-world universal speech enhancement.
    Sec. 3.1 relies entirely on challenge-provided data and the official blind set; external generalization is assumed, not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TS-URGENet: A Three-stage Universal Robust and Generalizable Speech Enhancement Network." pith.science (2026). https://pith.science/paper/T3QDSTFQ

@misc{pith2026250518533,
  author       = {Pith},
  title        = {Pith review of: TS-URGENet: A Three-stage Universal Robust and Generalizable Speech Enhancement Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3QDSTFQ}},
  note         = {Machine review of arXiv:2505.18533}
}
read the original abstract

Universal speech enhancement aims to handle input speech with different distortions and input formats. To tackle this challenge, we present TS-URGENet, a Three-Stage Universal, Robust, and Generalizable speech Enhancement Network. To address various distortions, the proposed system employs a novel three-stage architecture consisting of a filling stage, a separation stage, and a restoration stage. The filling stage mitigates packet loss by preliminarily filling lost regions under noise interference, ensuring signal continuity. The separation stage suppresses noise, reverberation, and clipping distortion to improve speech clarity. Finally, the restoration stage compensates for bandwidth limitation, codec artifacts, and residual packet loss distortion, refining the overall speech quality. Our proposed TS-URGENet achieved outstanding performance in the Interspeech 2025 URGENT Challenge, ranking 2nd in Track 1.

Figures

Figures reproduced from arXiv: 2505.18533 by the authors.

Figure 1
Figure 1. Overview of the TS-URGENet framework. 2. Method As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    restoration and enhancement

    Introduction Conventional speech enhancement (SE) primarily focuses on eliminating undesired signals such as noise and reverberation. In contrast, a more general SE task proposed in the Speech Signal Improvement (SSI) Challenges aims to improve speech quality by considering a broader range of factors, including col- oration, discontinuity, and loudness, i...

    arXiv 2025

  2. [2]

    1, our proposed framework is composed of three parts: filling, separation, and restoration stages

    Method As shown in Fig. 1, our proposed framework is composed of three parts: filling, separation, and restoration stages. We first employ a small-scale TF-GridNet [17] (TF-GridNet-S) as the filling module to fill the gaps caused by packet loss distortion. A residual connection is added after the filling stage to mini- mize information loss. Subsequently,...

    work page

  3. [3]

    Datasets We use the data provided in Track 1 of the Interspeech 2025 URGENT Challenge and employ the official scripts 5 for simu- lation

    Experiments 3.1. Datasets We use the data provided in Track 1 of the Interspeech 2025 URGENT Challenge and employ the official scripts 5 for simu- lation. The training data are dynamically simulated during the training process. The specific configuration is as follows: • Filling stage: We use all data with a sampling rate of at least 16 kHz, with all samp...

    work page 2025

  4. [4]

    6https://urgent-challenge.github.io/ urgent2025/rules/ Table 3: Results of the ablation study on the official validation set

    Results and Analysis In this section, we evaluate the proposed system using the offi- cially specified metrics6. 6https://urgent-challenge.github.io/ urgent2025/rules/ Table 3: Results of the ablation study on the official validation set. IDs Methods DNSMOS↑ NISQA↑ UTMOS↑ PESQ↑ ESTOI[28]↑ SDR (dB)↑ MCD↓ LSD↓ SpeechBERTScore↑ PhnSim↑ SpkSim↑ CAcc (%)↑ - No...

    work page 1911

  5. [5]

    We introduce a three-stage framework capa- ble of handling input speech with different distortions and sam- pling frequencies

    Conclusion This paper presents our submission to the Interspeech 2025 UR- GENT Challenge. We introduce a three-stage framework capa- ble of handling input speech with different distortions and sam- pling frequencies. Additionally, we propose an effective metric- aware fine-tuning function to boost various evaluation metrics specified in the challenge. Our...

    work page 2025

  6. [6]

    12274221) and the Yangtze River Delta Science and Technology Innovation Community Joint Re- search Project (Grant No

    Acknowledgment This work was supported by the National Natural Science Foun- dation of China (Grant No. 12274221) and the Yangtze River Delta Science and Technology Innovation Community Joint Re- search Project (Grant No. 2024CSJGG1100)

    work page

  7. [7]

    Audio codec enhancement with genera- tive adversarial networks,

    A. Biswas and D. Jia, “Audio codec enhancement with genera- tive adversarial networks,” in 2020 IEEE International Confer- ence on Acoustics, Speech and Signal Processing, ICASSP 2020, Barcelona, Spain, May 4-8, 2020. IEEE, 2020, pp. 356–360

    work page 2020

  8. [8]

    Icassp 2023 speech signal improvement challenge,

    R. Cutler, A. Saabas, B. Naderi, N.-C. Ristea, S. Braun, and S. Branets, “Icassp 2023 speech signal improvement challenge,” IEEE Open Journal of Signal Processing, 2024

    work page 2023

Show all 36 references
  1. [9]

    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, 2025

    2024

  2. [10]

    To- ward universal speech enhancement for diverse input conditions,

    W. Zhang, K. Saijo, Z.-Q. Wang, S. Watanabe, and Y . Qian, “To- ward universal speech enhancement for diverse input conditions,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–6

    2023

  3. [11]

    Toward universal speech enhancement for diverse input conditions,

    ——, “Toward universal speech enhancement for diverse input conditions,” in2023 IEEE Automatic Speech Recognition and Un- derstanding Workshop (ASRU). IEEE, 2023, pp. 1–6

    2023

  4. [12]

    Declipping speech using deep filter- ing,

    W. Mack and E. A. Habets, “Declipping speech using deep filter- ing,” in 2019 IEEE Workshop on Applications of Signal Process- ing to Audio and Acoustics (WASPAA). IEEE, 2019, pp. 200–204

    2019

  5. [13]

    Towards high-quality and efficient speech bandwidth extension with parallel amplitude and phase prediction,

    Y .-X. Lu, Y . Ai, H.-P. Du, and Z.-H. Ling, “Towards high-quality and efficient speech bandwidth extension with parallel amplitude and phase prediction,”ArXiv preprint, vol. abs/2401.06387, 2024

    2024 arXiv

  6. [14]

    Ks-net: Multi-band joint speech restoration and enhancement network for 2024 icassp ssi chal- lenge,

    G. Yu, R. Han, C. Xu, H. Zhao, N. Li, C. Zhang, X. Zheng, C. Zhou, Q. Huang, and B. Yu, “Ks-net: Multi-band joint speech restoration and enhancement network for 2024 icassp ssi chal- lenge,” in 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Work...

    2024

  7. [15]

    Bs- plcnet: Band-split packet loss concealment network with multi- task learning framework and multi-discriminators,

    Z. Zhang, J. Sun, X. Xia, C. Huang, Y . Xiao, and L. Xie, “Bs- plcnet: Band-split packet loss concealment network with multi- task learning framework and multi-discriminators,” in 2024 IEEE International Conference on Acoustics, Speech, and Signal Pro- cessing Workshops (ICASS...

    2024

  8. [16]

    V oicefixer: A unified framework for high-fidelity speech restoration,

    H. Liu, X. Liu, Q. Kong, Q. Tian, Y . Zhao, D. Wang, C. Huang, and Y . Wang, “V oicefixer: A unified framework for high-fidelity speech restoration,” in Interspeech 2022, 2022, pp. 4232–4236

    2022

  9. [17]

    Decoupling magnitude and phase estimation with deep resunet for music source separation,

    Q. Kong, Y . Cao, H. Liu, K. Choi, and Y . Wang, “Decoupling magnitude and phase estimation with deep resunet for music source separation,” in Proceedings of the 22nd International So- ciety for Music Information Retrieval Conference. ISMIR, Nov. 2021, pp. 342–349

    2021

  10. [18]

    Universal speech enhancement with score-based diffusion,

    J. Serr `a, S. Pascual, J. Pons, R. O. Araz, and D. Scaini, “Universal speech enhancement with score-based diffusion,” ArXiv preprint, vol. abs/2206.03065, 2022

    2022 arXiv

  11. [19]

    We employ the same training loss as in Gesper [13], consisting of reconstruction, adversarial, and feature-matching terms, with respective weights of 20, 1, and 1

    are adopted together. We employ the same training loss as in Gesper [13], consisting of reconstruction, adversarial, and feature-matching terms, with respective weights of 20, 1, and 1. 2.2. Separation stage Inspired by [5], we group clipping, noise, reverberation, and wind no...

  12. [20]

    Cascaded time + time-frequency unet for speech enhancement: Jointly addressing clipping, codec distortions, and gaps,

    A. A. Nair and K. Koishida, “Cascaded time + time-frequency unet for speech enhancement: Jointly addressing clipping, codec distortions, and gaps,” in IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2021, Toronto, ON, Canada, June 6-11, 2021. I...

    2021

  13. [21]

    Gesper: A unified framework for gen- eral speech restoration,

    J. Chen, Y . Shi, W. Liu, W. Rao, S. He, A. Li, Y . Wang, Z. Wu, S. Shang, and C. Zheng, “Gesper: A unified framework for gen- eral speech restoration,” in IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2...

    2023

  14. [22]

    Rad-net: A repairing and denoising network for speech signal improvement,

    M. Liu, Z. Chen, X. Yan, Y . Lv, X. Xia, C. Huang, Y . Xiao, and L. Xie, “Rad-net: A repairing and denoising network for speech signal improvement,” in 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), 2024, pp. 49–50

    2024

  15. [23]

    General speech restoration using two-stage generative adversarial networks,

    Q. Hu, T. Tan, M. Tang, Y . Hu, C. Zhu, and J. Lu, “General speech restoration using two-stage generative adversarial networks,” in 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 2024, pp. 31– 32

    2024

  16. [24]

    Tf-gridnet: Integrating full-and sub-band modeling for speech separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee et al. , “Tf-gridnet: Integrating full-and sub-band modeling for speech separation,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, 2023

    2023

  17. [25]

    Channel-wise subband input for better voice and accompaniment separation on high resolution music,

    H. Liu, L. Xie, J. Wu, and G. Yang, “Channel-wise subband input for better voice and accompaniment separation on high resolution music,” in Interspeech 2020. ISCA, 2020, pp. 1241–1245

    2020

  18. [26]

    Bigv- gan: A universal neural vocoder with large-scale training,

    S. Lee, W. Ping, B. Ginsburg, B. Catanzaro, and S. Yoon, “Bigv- gan: A universal neural vocoder with large-scale training,” in The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023

    2023

  19. [27]

    Mel-cepstral distance measure for objective speech quality assessment,

    R. Kubichek, “Mel-cepstral distance measure for objective speech quality assessment,” in Proceedings of IEEE pacific rim confer- ence on communications computers and signal processing, vol. 1. IEEE, 1993, pp. 125–128

    1993

  20. [28]

    Perceptual evaluation of speech quality (pesq): An objective method for end-to-end speech quality as- sessment of narrow-band telephone networks and speech codecs,

    I.-T. Recommendation, “Perceptual evaluation of speech quality (pesq): An objective method for end-to-end speech quality as- sessment of narrow-band telephone networks and speech codecs,” Rec. ITU-T P . 862, 2001

    2001

  21. [29]

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

    C. K. A. Reddy, V . Gopal, and R. Cutler, “Dnsmos P.835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in IEEE International Conference on Acous- tics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 202...

    2022

  22. [30]

    UTMOS: utokyo-sarulab system for voicemos challenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: utokyo-sarulab system for voicemos challenge 2022,” in Interspeech 2022. ISCA, 2022, pp. 4521– 4525

    2022

  23. [31]

    Evaluation metrics for generative speech en- hancement methods: Issues and perspectives,

    J. Pirklbauer, M. Sach, K. Fluyt, W. Tirry, W. Wardah, S. Moeller, and T. Fingscheidt, “Evaluation metrics for generative speech en- hancement methods: Issues and perspectives,” inSpeech Commu- nication; 15th ITG Conference. VDE, 2023, pp. 265–269

    2023

  24. [32]

    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 Inter- speech 2024, 2024, pp. 4943–4947

    2024

  25. [33]

    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. Xiaoet 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

  26. [34]

    Speechtok- enizer: Unified speech tokenizer for speech language models,

    X. Zhang, D. Zhang, S. Li, Y . Zhou, and X. Qiu, “Speechtok- enizer: Unified speech tokenizer for speech language models,” in The Twelfth International Conference on Learning Representa- tions, 2024

    2024

  27. [35]

    An algorithm for predicting the in- telligibility of speech masked by modulated noise maskers,

    J. Jensen and C. H. Taal, “An algorithm for predicting the in- telligibility of speech masked by modulated noise maskers,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 24, no. 11, pp. 2009–2022, 2016

    2009

  28. [36]

    Perceptual objective listening quality assessment (polqa), the third generation itu-t standard for end-to-end speech quality measurement part i—temporal align- ment,

    J. G. Beerends, C. Schmidmer, J. Berger, M. Obermann, R. Ull- mann, J. Pomy, and M. Keyhl, “Perceptual objective listening quality assessment (polqa), the third generation itu-t standard for end-to-end speech quality measurement part i—temporal align- ment,” Journal of the Aud...

    2013

Pith tools

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