Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

FlowSE: Efficient and High-Quality Speech Enhancement via Flow Matching

T0 review · 4 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 FlowSE claims that a flow-matching model with a Diffusion Transformer backbone outperforms existing generative speech enhancement methods while cutting inference latency more than tenfold.

desk verdict Plausible engineering advance, but the method section doesn't actually define the flow-matching loss and the equations contradict the architecture; not reproducible as written. read the letter →

arxiv 2505.19476 v2 pith:MWAHC22O submitted 2025-05-26 eess.AS eess.SP

Ziqian Wang , Zikai Liu , Xinfa Zhu , Yike Zhu , Mingshuai Liu , Jun Chen , Longshuai Xiao , Chao Weng
show 1 more author
Lei Xie
This is my paper
classification eess.ASeess.SP
keywords speechenhancementflowmatchingdiffusiontransformergenerativemodelmel-spectrogramtextconditioningreal-timefactorDNSMOS
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 introduces FlowSE, a speech enhancement model built on rectified flow matching. Instead of the iterative stochastic denoising used by diffusion models or the discrete-token quantization used by language-model approaches, FlowSE learns a single-pass velocity field that transports a noisy mel-spectrogram toward a clean one. The authors claim this yields higher DNSMOS perceptual quality and speaker similarity than diffusion and language-model baselines on the DNS Challenge test set, while running more than ten times faster than diffusion systems and reducing word error rate to about 8.8 percent. A key practical consequence is that generative speech enhancement can become fast enough for near-real-time use without sacrificing the quality advantages of generative modeling.

What carries the argument

The central object is the learned velocity field $v_\theta(M_t, t, C)$, defined on an interpolated mel-spectrogram $M_t$ that linearly blends the clean target $M_x$ with Gaussian noise, conditioned on optional text $C$. The field is parameterized by a 22-layer latent Diffusion Transformer (DiT) with 16 attention heads and hidden dimension 1024, and it is trained with a conditional flow-matching loss plus an $\ell^1$ mel reconstruction loss. At inference, the same field is integrated numerically in one ODE pass, with text conditioning randomly dropped during training so the model works both with and without transcripts.

What would settle it

Check whether any DNS Challenge 2021 test utterances or their speakers appear in the training mixtures described in Section 4.1; if they do, rerun the Table 1 comparison on a strictly disjoint held-out set and see whether the DNSMOS and speaker-similarity gaps over the diffusion and language-model baselines persist.

Watch

Extended reading notes

Core claim

The paper's central claim is that rectified flow matching, parameterized by a latent Diffusion Transformer operating directly on mel-spectrograms, simultaneously solves the two main weaknesses of prior generative speech enhancement: the information loss from quantizing speech into discrete tokens, and the high inference cost of iterative diffusion sampling. FlowSE is trained with a conditional flow-matching loss together with an L1 mel-spectrogram reconstruction loss, using noisy mel-spectrograms and optional character-level transcripts as conditions. At inference, the learned velocity field is integrated with an ODE solver in a single pass, and a pre-trained vocoder converts the enhanced mel-spectrogram to a waveform. The paper reports the highest DNSMOS scores and speaker similarity on the DNS Challenge test set among compared systems, a real-time factor of 0.31 on a single NVIDIA 4090D, and a word error rate of about 8.8 percent on a simulated test set, versus 14 to 15 percent for diffusion baselines.

Load-bearing premise

The paper trains on DNS Challenge 2021 datasets and evaluates on the DNS Challenge 2021 test set without explicitly stating that the test set was excluded from training, so if the test utterances or speakers overlap the training data, the reported quality and speaker-similarity gains on that test set would be inflated.

Editorial extensions

If this is right

  • If the reported results hold, generative speech enhancement can operate in near real time on a single consumer GPU, removing the latency barrier that previously limited diffusion-based enhancement to offline use.
  • Text-based conditioning becomes an optional plug-in that improves enhancement quality when transcripts are available, without requiring the model to be retrained for the text-free case.
  • The avoidance of discrete token quantization means speaker identity and intelligibility do not degrade from codec artifacts, a weakness the paper attributes to language-model-based enhancement.
  • The single-pass noisy-to-clean transport could generalize to other speech restoration tasks such as dereverberation, bandwidth extension, and historical recording restoration.

Reading between the lines

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

  • Assuming the DNS test set is genuinely disjoint from the training data, flow matching appears to offer a favorable quality-speed trade-off that could make it the default generative backend for latency-sensitive speech enhancement products.
  • The text-conditioning gain is likely concentrated in low-SNR conditions where phonetic context matters most; an SNR-stratified evaluation of DNSMOS and WER would make this explicit.
  • The reported real-time factor bundles the pre-trained Vocos vocoder, so absolute RTF depends on vocoder choice, though the relative advantage over iterative diffusion would persist.
  • The same rectified-flow formulation, with minimal changes, could be applied to audio super-resolution or target-speaker extraction, where a continuous transport from degraded to clean acoustic space is also natural.
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 / 5 minor

Summary. FlowSE proposes a flow-matching-based speech enhancement system that operates on mel-spectrograms with optional text conditioning. The paper describes a DiT-based velocity model, a training setup using an L1 mel-spectrogram loss, and evaluations on the DNS Challenge and a simulated test set, reporting DNSMOS, speaker similarity, WER, and RTF comparisons against diffusion and language-model baselines. The central claim is that FlowSE significantly outperforms prior generative methods while being substantially faster. The manuscript also states that code, pre-trained checkpoints, and audio samples will be released.

Significance. The target task is important, and the combination of rectified flow matching with a DiT backbone directly on mel-spectrograms, with optional text conditioning, is a plausible and potentially useful direction. The paper includes comparisons of multiple baselines with the same vocoder, reports both quality and efficiency metrics, and promises public code and checkpoints. However, the technical specification contains serious gaps: the flow-matching objective is never actually stated, and the ODE endpoints in Section 3.2 conflict with the training and inference description in Section 3.3.2. These are load-bearing issues that prevent the reported numbers from being traced to a reproducible training procedure. If the method is properly specified and the dataset overlap concern is resolved, the work could be a meaningful contribution; in its current form, the central claim is not supported.

major comments (4)
  1. [3.2, Eq. (1); 3.3.2] Equation (1) defines the ODE with z_0 = y and z_1 = x, so the learned velocity should transport noisy speech to clean speech. Section 3.3.2 instead states that training uses an interpolated mel-spectrogram M_t obtained by linearly combining the clean target M_x with Gaussian noise, and that inference starts from Gaussian noise M_0, with M_y used only as a conditioning input. These are two different generative processes. If inference starts from M_0, Eq. (1) has the wrong initial condition; if it starts from M_y, the training interpolation is inconsistent with the ODE. The authors must specify the actual forward process and align Eq. (1) with Eqs. (3)-(4) and the training interpolation.
  2. [3.4, Eq. (7)] The 'Training Objective' section never states the conditional flow-matching loss mentioned in the abstract and Section 3.2. Equation (7) is an L1 reconstruction loss on mel-spectrograms, not a regression toward a target velocity field. Without an equation of the form L = E[||v_theta(M_t, t, C) - u_t(M_x, M_0, M_y)||^2] with an explicit target velocity u_t and interpolation schedule, the training procedure is under-specified. This is not a stylistic issue: the reported DNSMOS, WER, and RTF numbers cannot be traced to a well-defined method without this equation.
  3. [4.1] The training set description lists the DNS Challenge - Interspeech 2021 datasets, and the test evaluation uses the DNS Challenge - Interspeech 2021 test set. The paper never states that the test set was held out from the training mixture. If any test utterances or noise conditions overlap the training data, the DNSMOS and speaker-similarity improvements in Table 1 would be inflated. The authors should clarify the exact data split and explicitly confirm that no test material was used in training.
  4. [4.4, Table 2] The inference procedure is not fully specified: the number of ODE solver steps used for the RTF and WER measurements is not reported. Flow-matching performance and latency depend strongly on this number, so the claimed RTF advantage of 0.31 versus diffusion baselines above 3.0 is not verifiable without this information. Please report the solver type, number of steps, and any early-stopping criteria used at inference.
minor comments (5)
  1. [3.3.2] The text calls the backbone a 'latent Diffusion Transformer' but says the model operates directly on the mel-spectrogram domain; please clarify whether a latent space is actually used or whether 'latent' is a naming convention.
  2. [3.3.3, Eq. (6)] Equation (6) uses M_x for the enhanced mel-spectrogram, but M_x denotes the clean target elsewhere; use a distinct symbol such as \hat{M}_x to avoid confusion.
  3. [4.1] The simulated test set is described only as mixing VCTK speech with unseen WHAM! and DEMAND noise at SNRs from -5 to 10 dB; the number of utterances, exact SNR distribution, and whether the noise is seen in training should be stated for reproducibility.
  4. [4.4] The DNSMOS and speaker-similarity differences between FlowSE-w/ text and FlowSE-w/o text are extremely small (e.g., 3.690 vs. 3.685 in one condition); reporting confidence intervals or significance tests would help assess whether the text-conditioned variant is genuinely better.
  5. [2.2] The phrase 'first introduced by Meta' is informal; please cite Lipman et al. directly in the text rather than by company name. Also fix the typo 'optimizing a condition flow matching loss' in the abstract.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: no reported quantity reduces to a fitted input, though the paper has minor non-load-bearing self-citations and two non-circular validity risks (unspecified flow-matching loss; DNS train/test overlap).

full rationale

FlowSE's central claims (Tables 1 and 2) are benchmark comparisons against external systems using DNSMOS, WeSpeaker cosine similarity, and Whisper WER; none of these metrics is a fitted parameter of FlowSE, and no equation in Section 3 turns a reported prediction into a training target by construction. The self-citations to SELM [1] and DCCRN [18] share authors with the present paper, but they are used as baselines or related-work context only and are not load-bearing for FlowSE's design or results; this is the minor self-citation that keeps the score at 2 rather than 0. The paper does omit the actual conditional flow-matching objective: Section 3.4 provides only the L1 mel loss (Eq. 7), and the ODE in Eq. (1) (z0=y, z1=x) conflicts with the training interpolation and inference initialization described in Section 3.3.2, where M_t mixes clean M_x with Gaussian noise and inference starts from M_0. That is an internal-consistency and reproducibility gap, not a circular reduction. Section 4.1 lists the DNS Challenge 2021 dataset in both the training mixture and the test set without an explicit holdout statement, which is a potential data-leakage risk for the DNSMOS and Spk Sim results, but this is not an equation-level circularity. The use of Whisper both to generate training transcripts and to score WER is an evaluation-bias concern, not a fitted-input prediction. Overall, no load-bearing circular step is exhibited.

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

The central claim depends on several unstated choices and assumptions. Most importantly, the flow matching training loss is never defined, the interpolation schedule is unspecified, and the question of DNS train/test overlap is left open. These are the main burdens beyond the prior literature.

free parameters (4)
  • Inference ODE steps = not reported
    The number of ODE solver steps controls the trade-off between output quality and the reported RTF of 0.31; without this number the speed claim is not reproducible.
  • Text dropout probability = not reported
    Used during training to make the model work without text at inference; the value affects the balance between text-conditioned and text-free performance.
  • Training data mixture ratios = not reported
    Proportions of WeNetSpeech, GigaSpeech, VCTK, and DNS data affect model robustness and may influence test results.
  • Interpolation schedule for z_t = not reported
    The paper says M_t is a linear combination of clean M_x and Gaussian noise, but the exact interpolation formula is not given, and it determines the velocity target and the entire flow matching objective.
assumptions (6)
  • standard math Flow matching learns a velocity field that transports a source distribution to a target distribution (Lipman et al., 2023)
    The model relies on the theoretical guarantee that optimizing a flow matching objective yields a valid probability flow.
  • domain assumption Mel spectrograms retain enough information for high-quality waveform reconstruction via a vocoder
    The system uses Vocos to convert predicted mel spectrograms to waveforms; if information is lost in the mel representation, output quality would be capped.
  • domain assumption DNSMOS, WeSpeaker cosine similarity, and Whisper WER are valid proxies for human perceptual quality, speaker identity, and intelligibility
    All headline claims rest on these reference-free or ASR-based metrics rather than human listening tests.
  • domain assumption The DNS Challenge 2021 test set was not used during training
    The paper lists DNS 2021 datasets in the training mixture and evaluates on the DNS 2021 test set, but never states the test set was held out.
  • domain assumption Whisper transcriptions used for training text conditioning are accurate enough to be useful
    If the auto-generated transcripts contain systematic errors, the text-conditioned variant could learn speech-to-incorrect-text alignments.
  • ad hoc to paper Randomly dropping text during training yields a model that works both with and without text at inference
    This is a design choice stated without ablations or evidence of robustness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowSE: Efficient and High-Quality Speech Enhancement via Flow Matching." pith.science (2026). https://pith.science/paper/MWAHC22O

@misc{pith2026250519476,
  author       = {Pith},
  title        = {Pith review of: FlowSE: Efficient and High-Quality Speech Enhancement via Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MWAHC22O}},
  note         = {Machine review of arXiv:2505.19476}
}
read the original abstract

Generative models have excelled in audio tasks using approaches such as language models, diffusion, and flow matching. However, existing generative approaches for speech enhancement (SE) face notable challenges: language model-based methods suffer from quantization loss, leading to compromised speaker similarity and intelligibility, while diffusion models require complex training and high inference latency. To address these challenges, we propose FlowSE, a flow-matching-based model for SE. Flow matching learns a continuous transformation between noisy and clean speech distributions in a single pass, significantly reducing inference latency while maintaining high-quality reconstruction. Specifically, FlowSE trains on noisy mel spectrograms and optional character sequences, optimizing a conditional flow matching loss with ground-truth mel spectrograms as supervision. It implicitly learns speech's temporal-spectral structure and text-speech alignment. During inference, FlowSE can operate with or without textual information, achieving impressive results in both scenarios, with further improvements when transcripts are available. Extensive experiments demonstrate that FlowSE significantly outperforms state-of-the-art generative methods, establishing a new paradigm for generative-based SE and demonstrating the potential of flow matching to advance the field. Our code, pre-trained checkpoints, and audio samples are available.

Figures

Figures reproduced from arXiv: 2505.19476 by the authors.

Figure 1
Figure 1. Overview of FlowSE. During training (left), the model takes a noisy mel-spectrogram My, an interpolated mel￾spectrogram Mt (a mixture of clean speech and Gaussian noise), and an optional transcript C. The text is processed by a text encoder T and concatenated with the audio embed￾dings. The DiT-based flow model F predicts the probability flow, which is trained via gradient updates to reconstruct the clean speech mel… view at source ↗

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

    eess.AS 2026-07 conditional novelty 6.0 of 10

    Magnitude strength, not estimated phase, drives SE-induced ASR degradation, and the optimal strength is recognizer-dependent (strong for wav2vec 2.0, mild for Whisper).

  2. UniFlow: Unifying Speech Front-End Tasks via Continuous Generative Modeling

    eess.AS 2025-08 conditional novelty 6.0 of 10

    UniFlow unifies four speech front-end tasks in one continuous-latent generative model with task-ID conditioning and reports competitive, but not uniformly superior, benchmark scores.

Reference graph

Works this paper leans on

41 extracted references · 18 canonical work pages · cited by 2 Pith papers

  1. [1]

    While traditional deterministic methods can attenuate noise, they often struggle to preserve speech naturalness under challenging conditions

    Introduction Speech enhancement (SE) aims to recover clean speech from noisy signals, playing a vital role in applications such as telecommunications, hearing aids, and speech recognition front ends. While traditional deterministic methods can attenuate noise, they often struggle to preserve speech naturalness under challenging conditions. Recent advances...

    work page

  2. [2]

    Related Work 2.1. Generative-Based Speech Enhancement Generative approaches have recently become a prominent focus in speech enhancement (SE), effectively restoring clean speech signals from noisy inputs. Traditional deep learning-based SE methods, such as convolutional neural networks (CNNs) and re- current neural networks (RNNs), primarily focus on dete...

    arXiv 2025

  3. [3]

    Proposed Approach 3.1. Overall Framework Speech enhancement (SE) aims to recover clean speechx∈R T from its noisy observationy∈R T , a task that requires effective modeling of both acoustic signals and any available auxiliary information. To address this challenge, we propose FlowSE, a novel approach that integrates a flow-matching framework to achieve ef...

    work page

  4. [4]

    Experiments 4.1. Datasets & Evaluation Metrics Training setsTo comprehensively evaluate the effective- ness of FlowSE, we construct a large-scale training dataset by combining multiple publicly available speech and noise datasets. We utilize WeNetSpeech [25], GigaSpeech [26], V oiceBank(VCTK) [27], and the DNS Challenge - Interspeech 2021 datasets [28] to...

    work page arXiv 2021

  5. [5]

    FlowSE efficiently reconstructs high-quality speech while preserving speaker characteristics, achieving state-of-the-art performance

    Conclusion In this work, we introduce FlowSE, a novel speech enhance- ment model based on flow matching, which addresses the lim- itations of diffusion models’ high complexity and slow in- ference, as well as the quantization loss in language-model- based approaches. FlowSE efficiently reconstructs high-quality speech while preserving speaker characterist...

    work page

  6. [6]

    Selm: Speech enhancement using discrete tokens and language models,

    Z. Wang, X. Zhu, Z. Zhang, Y . Lv, N. Jiang, G. Zhao, and L. Xie, “Selm: Speech enhancement using discrete tokens and language models,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 561–11 565

    work page 2024

  7. [7]

    Masksr: Masked language model for full-band speech restoration,

    X. Li, Q. Wang, and X. Liu, “Masksr: Masked language model for full-band speech restoration,”arXiv preprint arXiv:2406.02092, 2024

    arXiv 2024

  8. [8]

    Neural discrete represen- tation learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete represen- tation learning,”Advances in neural information processing sys- tems, vol. 30, 2017

    2017

Show all 41 references
  1. [9]

    Soundstream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “Soundstream: An end-to-end neural audio codec,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 30, pp. 495–507, 2021

    2021

  2. [10]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,”arXiv preprint arXiv:2210.13438, 2022

    2022 arXiv

  3. [11]

    Conditional diffusion probabilistic model for speech en- hancement,

    Y .-J. Lu, Z.-Q. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional diffusion probabilistic model for speech en- hancement,” inProc. ICASSP. IEEE, 2022, pp. 7402–7406

    2022

  4. [12]

    Speech enhancement with score-based generative models in the complex stft domain,

    S. Welker, J. Richter, and T. Gerkmann, “Speech enhancement with score-based generative models in the complex stft domain,” arXiv preprint arXiv:2203.17004, 2022

    2022 arXiv

  5. [13]

    Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,

    J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, vol. 31, pp. 2724–2737, 2023

    2023

  6. [14]

    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,

  7. [15]

    Flowsep: Language-queried sound separation with rectified flow matching,

    Y . Yuan, X. Liu, H. Liu, M. D. Plumbley, and W. Wang, “Flowsep: Language-queried sound separation with rectified flow matching,” inICASSP 2025-2025 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  8. [16]

    Matcha-tts: A fast tts architecture with conditional flow match- ing,

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha-tts: A fast tts architecture with conditional flow match- ing,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 341–11 345

    2024

  9. [17]

    E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,

    S. E. Eskimez, X. Wang, M. Thakker, C. Li, C.-H. Tsai, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tanet al., “E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 682– 689

    2024

  10. [18]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  11. [19]

    A flow-based deep latent variable model for speech spectrogram modeling and en- hancement,

    A. A. Nugraha, K. Sekiguchi, and K. Yoshii, “A flow-based deep latent variable model for speech spectrogram modeling and en- hancement,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 28, pp. 1104–1117, 2020

    2020

  12. [20]

    Flowavse: Efficient audio-visual speech enhancement with conditional flow match- ing,

    C. Jung, S. Lee, J.-H. Kim, and J. S. Chung, “Flowavse: Efficient audio-visual speech enhancement with conditional flow match- ing,”arXiv preprint arXiv:2406.09286, 2024

    2024 arXiv

  13. [21]

    Speech enhancement and dereverberation with diffusion-based generative models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech enhancement and dereverberation with diffusion-based generative models,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2351–2364, 2023

    2023

  14. [22]

    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, no. 8, pp. 1256–1266, 2019

    2019

  15. [23]

    Real time speech enhancement in the waveform domain,

    A. Defossez, G. Synnaeve, and Y . Adi, “Real time speech enhancement in the waveform domain,”arXiv preprint arXiv:2006.12847, 2020

    2006 arXiv

  16. [24]

    Dccrn: Deep complex convolution re- current 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 re- current network for phase-aware speech enhancement,”arXiv preprint arXiv:2008.00264, 2020

    2008 arXiv

  17. [25]

    Cp-gan: Context pyra- mid generative adversarial network for speech enhancement,

    G. Liu, K. Gong, X. Liang, and Z. Chen, “Cp-gan: Context pyra- mid generative adversarial network for speech enhancement,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 6624– 6628

    2020

  18. [26]

    A multi-resolution approach to gan-based speech enhancement,

    H. Y . Kim, J. W. Yoon, S. J. Cheon, W. H. Kang, and N. S. Kim, “A multi-resolution approach to gan-based speech enhancement,” Applied Sciences, vol. 11, no. 2, p. 721, 2021

    2021

  19. [27]

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

    C. Veaux, J. Yamagishi, and S. King, “The voice bank corpus: De- sign, collection and data analysis of a large regional accent speech database,” 11 2013, pp. 1–4

    2013

  20. [28]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  21. [29]

    V ocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis,

    H. Siuzdak, “V ocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis,” arXiv preprint arXiv:2306.00814, 2023

    2023 arXiv

  22. [30]

    Bigvgan: A universal neural vocoder with large-scale training,

    S.-g. Lee, W. Ping, B. Ginsburg, B. Catanzaro, and S. Yoon, “Bigvgan: A universal neural vocoder with large-scale training,” arXiv preprint arXiv:2206.04658, 2022

    2022 arXiv

  23. [31]

    Wenet: Production oriented streaming and non- streaming end-to-end speech recognition toolkit,

    Z. Yaoet al., “Wenet: Production oriented streaming and non- streaming end-to-end speech recognition toolkit,” inProc. Inter- speech. ISCA, 2021, pp. 4054–4058

    2021

  24. [32]

    GigaSpeech: An Evolving, Multi-Domain ASR Corpus with 10,000 Hours of Transcribed Audio,

    G. Chenet al., “GigaSpeech: An Evolving, Multi-Domain ASR Corpus with 10,000 Hours of Transcribed Audio,” inProc. Inter- speech, 2021, pp. 3670–3674

    2021

  25. [33]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  26. [34]

    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

  27. [35]

    Wham!: Extend- ing speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. L. Roux, “Wham!: Extend- ing speech separation to noisy environments,”arXiv preprint arXiv:1907.01160, 2019

    1907 arXiv

  28. [36]

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

    J. Thiemann, N. Ito, and E. Vincent, “The diverse environments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings,” inProceedings of Meetings on Acoustics, vol. 19, no. 1. AIP Publishing, 2013

    2013

  29. [37]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” inProc. ICASSP. IEEE, 2017, pp. 5220– 5224

    2017

  30. [38]

    Robust speech recognition via large- scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large- scale weak supervision,” 2022. [Online]. Available: https: //arxiv.org/abs/2212.04356

    2022 arXiv

  31. [40]

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

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watan- abe, “Tf-gridnet: Integrating full-and sub-band modeling for speech separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    2023

  32. [41]

    Convnext v2: Co-designing and scaling convnets with masked autoencoders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, and S. Xie, “Convnext v2: Co-designing and scaling convnets with masked autoencoders,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2023, pp. 16 133–16 142

    2023

  33. [2023]

    Available: https://openreview.net/forum?id= PqvMRDCJT9t

    [Online]. Available: https://openreview.net/forum?id= PqvMRDCJT9t

Pith tools

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