Pith. sign in

REVIEW 4 major objections 4 minor 35 references

A pretrained clean-speech diffusion model can be steered into a speech enhancer by the gradient of a small noise model trained on a short noise clip.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

GDiffuSE steers a frozen speech-diffusion generator toward clean speech using a lightweight noise-model likelihood adapted from a short reference noise clip, improving PESQ/SI-SDR on mismatched BBC noise.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The core idea—guiding a frozen clean-speech diffusion model with a lightweight per-noise density model—is genuinely new and worth discussion, but the paper's central derivation has a load-bearing error and the abstract's 'consistent improvements' is only true for PESQ/SI-SDR, not for STOI/DNSMOS. Needs a corrected derivation and a more careful empirical framing. the 4 major comments →

arxiv 2510.04157 v2 pith:MFV5KXJY submitted 2025-10-05 cs.SD eess.AS

GDiffuSE: Diffusion-based speech enhancement with noise model guidance

Efrayim Yanir , David Burshtein , Sharon Gannot This is my paper
classification cs.SD eess.AS
keywords diffusion probabilistic modelsspeech enhancementDDPM guidancenoise modelgenerative speech enhancementcausal convolutional networkunconditional diffusion priorreverse sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

GDiffuSE claims that speech enhancement can be reformulated as guided generation: keep a diffusion model trained only on clean speech frozen, train a compact per-step causal network to model the distribution of the combined acoustic-plus-diffusion noise from a short reference noise clip, and use that network's log-density gradient to steer each reverse step of the diffusion toward the clean signal. The paper argues this is a lighter and more transferable route than retraining a heavy denoiser for each noise type, because adapting to a new noise only requires training the small noise model. On noisy recordings made by mixing read speech with out-of-domain sound-effects noise, the paper reports consistently higher objective speech-quality and signal-to-distortion scores than a state-of-the-art score-based generative baseline across SNR levels, with larger gains on high-frequency noise. If correct, the result implies that one foundation speech generator can serve many acoustic conditions by swapping a lightweight noise model.

Core claim

The paper's central claim is that DDPM guidance can be driven by the noise itself rather than by a classifier on the clean signal. Writing the noisy observation y = x0 + w through the diffusion marginal gives y = (1/sqrt(ᾱ_t)) x_t - g(t)ê_t + w, so the likelihood of y given the current diffusion state x_t is the density of v_t = w - g(t)ê_t. The paper trains a lightweight causal convolutional network, one model per diffusion step, to approximate that density from samples synthesized from a reference noise clip, and then adds the gradient of its log-density to the reverse-diffusion mean with a scale that grows as the effective SNR rises. The pretrained clean-speech diffusion backbone stays fr

What carries the argument

The load-bearing object is the guidance gradient ∇_x log p_phi(y|x) in Eq. (10), which is appended to the DDPM reverse mean with an SNR-dependent scale s_t. The density p_phi(y|x) is built from a family of causal convolutional noise models φ_t that predict the conditional mean and variance of v_t = w - g(t)ê_t, where ê_t is the forward-diffusion noise and g(t) is the diffusion noise level at step t (Eqs. (13)-(19)). The models are trained by maximum likelihood on samples v_t generated from a short noise reference clip; at inference, the same models evaluate the gradient of the loss with respect to the current noise estimate v_t = y - (1/sqrt(α_t)) μ_θ(x_t,t). This machinery converts a frozen

Load-bearing premise

The load-bearing premise is that the forward-diffusion noise ê_t is statistically independent of the diffusion state x_t, which lets p(y|x_t) be treated as the density of w - g(t)ê_t; because Eq. (4) makes ê_t a deterministic function of x_t and x0, the premise does not literally hold, and the method also assumes the true acoustic noise is stationary between the reference clip and the utterance being enhanced.

What would settle it

Take a fixed clean utterance x0 and a fixed noise w, compute the exact density p(y|x_t) implied by Eq. (4) together with y = x0 + w, and compare its gradient with the paper's guidance gradient from Eq. (10) and Algorithm 2. If the two gradients point in materially different directions, the guided sampler is not sampling from the intended conditional distribution. A second check: run Algorithm 2 with v_t computed as in the paper and with v_t computed from an explicit x0 estimate, (x_t - sqrt(1-ᾱ_t)εθ)/sqrt(ᾱ_t), and measure SI-SDR; if results are indistinguishable, the reported gains come from

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A new noise condition can be handled by training only a compact per-step noise model on a short noise reference clip, leaving the large clean-speech diffusion model untouched.
  • The reported gains in objective speech quality and signal-to-distortion over a strong generative baseline, if they hold, mean a single foundation speech generator can be repurposed for enhancement across unfamiliar noise types.
  • Because only the noise model is trained, the approach sidesteps collecting large matched noisy-clean corpora for each new environment, provided noise statistics are stable between the reference clip and the test utterance.
  • The SNR-dependent guidance schedule implies that the enhancement is most heavily influenced by the noise model in the later, less noisy stages of the reverse process, where the noise-model estimate is most reliable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the independence assumption on the forward-diffusion noise is repaired, the same guidance scheme may become a principled posterior sampler; the current version would then be an approximation whose empirical success deserves a mechanism-level explanation rather than being taken as the exact Bayes update.
  • The general recipe—condition a frozen generative prior on an observation via the gradient of a learned degradation model—extends naturally to other additive degradation problems, such as dereverberation or restoration of old recordings, whenever a short reference of the degradation alone is available.
  • A natural stress-test is to expose the method to noise whose spectrum drifts within a single test utterance; the paper's own stability requirement predicts a smooth degradation as drift increases, and measuring that curve would tell practitioners how long a reference clip remains valid.
  • The discrepancy between the v_t formula in Algorithm 2 and the x0-estimate expression in the derivation is testable: re-running the sampler with the alternative v_t would show whether the reported gains are robust to that algebraic choice or hinge on it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes GDiffuSE, a speech-enhancement method that guides a pretrained unconditional clean-speech DDPM with a lightweight per-step noise model. The clean speech is generated by the frozen diffusion model, and a compact CNN noise model, trained on a short noise-only reference clip, supplies a guidance gradient that steers the reverse process toward the clean speech (Algorithms 1 and 2). The authors evaluate on LibriSpeech utterances mixed with BBC sound-effects noise and report that GDiffuSE outperforms SGMSE in PESQ and SI-SDR at all tested SNR levels, while having lower STOI and DNSMOS.

Significance. If the proposed guidance mechanism were sound, the paper would make a useful practical contribution: adapting a large pretrained speech-generation DDPM to unseen noise types by training only a small noise model, with no retraining of the backbone. The idea of using a learned noise density as a guidance signal is appealing, and the experimental setup (out-of-domain LibriSpeech clean speech, BBC noise unseen by SGMSE) is well chosen for demonstrating robustness. However, the central derivation in §3.2 is mathematically invalid, and the algorithm as implemented does not match the derived guidance gradient. Because the main claim rests on this derivation, the contribution is not established in its current form.

major comments (4)
  1. [§3.2, Eq. (12) versus Eqs. (13)-(15)] Equation (12) gives y = (1/√ᾱ_t)x_t − √((1−ᾱ_t)/ᾱ_t) ê_t + w. The combined noise v_t defined in Eq. (13) must therefore have coefficient √((1−ᾱ_t)/ᾱ_t), not √((1−α_t)/α_t). Yet Eqs. (13)-(14) define g(t) = √((1−α_t)/α_t), and Eq. (15) writes p(y|x_t) = p_Vt|Xt(y − (1/√α_t)x_t | x_t). This α_t/ᾱ_t conflation propagates to Algorithm 2, where the guidance input is computed with α_t instead of ᾱ_t. This is a load-bearing algebraic error: the density being modeled in training (Algorithm 1) is for a different random variable than the one used in the guidance update.
  2. [§3.2, independence assumption for ê_t] The derivation reduces p(y|x_t) to the marginal p_V(v_t) by assuming that ê_t is independent of x_t. This is false: Eq. (4) defines x_t = √ᾱ_t x_0 + √(1−ᾱ_t) ê_t, so ê_t is, for any fixed x_0, a deterministic function of x_t. Even after marginalizing over x_0, ê_t and x_t are dependent; in particular E[ê_t | x_t] is generally nonzero and depends on x_t. Therefore v_t = w − g(t)ê_t and x_t are correlated, and p(y|x_t) is not equal to p_V(y − (1/√ᾱ_t)x_t). The guidance gradient in Eq. (10) is consequently not ∇_{x_t} log p(y|x_t), as claimed. This invalidates the theoretical justification of the method.
  3. [Algorithm 2, lines 5 and 8] Even if the derivation were corrected, the implemented algorithm does not follow from it. Eq. (12) and Eq. (15) require v_t = y − (1/√ᾱ_t)x_t, but Algorithm 2 line 5 sets v_t ← y − (1/√α_t)μ_θ(x_t,t), replacing the current state x_t with the denoised posterior mean and using α_t instead of ᾱ_t. Similarly, line 8 uses the prefactor β_t/√α_t and the derivative −(1/√α_t)∂loss/∂v_t, whereas Eq. (12) would require the analogous quantity with √ᾱ_t and x_t. The algorithm therefore does not implement the conditional reverse update derived in §3.2. If the formula is corrected, the method changes materially, so the reported experiments do not test the proposed derivation.
  4. [§4.1, guidance hyperparameters] The guidance schedule's λ_max and γ are calibrated on one clip per SNR level and then used to report results on the same SNR levels. No separate validation set or cross-validation procedure is described, so the reported 'consistent improvements' may partially reflect selection on these clips. The authors should describe how the calibration clips were chosen, whether they overlap with the test set, and whether the reported numbers are averages over runs with fixed hyperparameters or selected after seeing test results. This is necessary to evaluate the empirical claim in Tables 1 and 2.
minor comments (4)
  1. [§4.2 and Tables 1-2] The abbreviations 'sgmseW' and 'sgmseT' are used in Table 1 without definition; the text says SGMSE was trained on WSJ0 or TIMIT, so the labels should be defined explicitly (e.g., SGMSE-WSJ0, SGMSE-TIMIT).
  2. [Throughout] Typographical errors: 'follo’ing' in §2, 'sprocesss' in §3.1, 'V AEs' in §1. The paper would benefit from a careful proofread.
  3. [§4.5] The claim in the abstract that GDiffuSE gives 'consistent improvements' is stronger than what Table 1 shows: SGMSE-W achieves higher STOI and DNSMOS at all SNR levels. The text should state precisely which metrics improve and acknowledge the trade-offs.
  4. [Eq. (10)] The gradient in Eq. (10) is evaluated at x = μ(x_t,t), but the conditional density in Eq. (15) is a function of x_t, not of μ(x_t,t). This distinction is not explained and contributes to the mismatch with Algorithm 2.

Circularity Check

0 steps flagged

No significant circularity: the guidance derivation follows standard classifier guidance and the reported metrics are held-out evaluations, not fitted quantities.

full rationale

The paper's central output is an enhanced waveform produced by a pretrained DDPM guided by a separately trained noise model; the reported PESQ/SI-SDR values are computed against held-out clean references and are not fitted quantities. The derivation of the guidance gradient (Eqs. 9–10) adopts the standard classifier-guidance form from Dhariwal & Nichol (2021), and the conditional density p_φ(y|x_t) is introduced through an explicitly stated modeling assumption that ê_t is independent of x_t. That assumption is mathematically inconsistent with Eq. (4) — a correctness concern — but it is not circular: the noise model is trained on synthetic v_t samples generated from a reference noise clip and then applied to different noise segments of the same type, so the guidance signal is not equivalent to the evaluation target by construction. The hyperparameters λ_max and γ are calibrated per SNR level and used at the same SNR levels, but the paper states this calibration explicitly and the test set comprises 20 speakers rather than the calibration clips, so no reported metric is forced by construction. The paper contains no load-bearing self-citations and does not import a uniqueness theorem from the authors' own prior work; references [5], [14], and [19] are external prior methods. The footnote in §4.5 acknowledging future work on characterizing noise types is a scope limitation, not evidence of circularity. Overall, the derivation chain contains no step in which a predicted quantity is equivalent to a fitted input or imported self-result.

Axiom & Free-Parameter Ledger

2 free parameters · 7 axioms · 0 invented entities

The paper adds one small learned module (per-step noise model) but no invented physical entities. It relies on standard DDPM background, a stated-but-false independence premise, the availability/stability of a reference noise clip, the expressiveness of a Gaussian autoregressive CNN for noise log-density, and the validity of the pretrained UnDiff speech prior. Two guidance hyperparameters are calibrated on data (one clip per SNR level).

free parameters (2)
  • guidance scale λ_max(SNR) = [0.8, 0.72, 0.6, 0.55] for SNR [10, 5, 0, -5] dB
    Chosen in §4.1 by calibrating on one clip per SNR level; directly scales the guidance gradient in Eq. (10)/Algorithm 2, so it has a first-order effect on the reported metrics.
  • guidance exponent γ = 0.7
    Hand-calibrated on one clip per SNR level (§4.1); controls how quickly guidance strength decays with diffusion noise level; interacts with λ_max.
axioms (7)
  • standard math Standard DDPM forward/reverse Markov-chain equations and noise-prediction parameterization (Eqs. (2)-(8), from [3,14]).
    Invoked without proof throughout §3.1; accepted background.
  • ad hoc to paper The diffusion noise ê_t is independent of the current state x_t (assumed in §3.2 to simplify p(y|x_t)).
    This is the paper's explicit simplification; it is inconsistent with Eq. (4), where ê_t is the noise that produced x_t.
  • domain assumption A noise-only reference segment w̄ from the same distribution as the acoustic noise w is available (VAD-segmentable) to train the noise models.
    Required by Algorithm 1; central to the adaptation claim.
  • domain assumption The noise statistics do not change between the reference clip used in training and the noise encountered at inference.
    Stated in the abstract and §3: 'provided that the noise statistics has not significantly changed between train and inference time'.
  • ad hoc to paper p(y|x_t) can be evaluated using the surrogate v_t = y - (1/√α_t)μ_θ(xt,t) and the noise model ϕ_t, i.e., guidance can be applied at the denoised posterior mean rather than the current state x_t.
    Algorithm 2 lines 5-8; not derived from Eq. (12)-(15), which use x_t with ᾱ_t.
  • domain assumption The pretrained UnDiff model trained on VCTK/LJSpeech clean speech provides a valid speech prior for LibriSpeech test utterances.
    No re-training or domain-adaptation of the backbone is performed, so its prior is taken as given.
  • domain assumption The per-sample conditional Gaussian parameterization in Eq. (18) is sufficiently expressive to represent the combined noise density w - g(t)ê_t.
    All guidance quality depends on this Gaussian autoregressive approximation.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GDiffuSE: Diffusion-based speech enhancement with noise model guidance." pith.science (2026). https://pith.science/paper/MFV5KXJY

@misc{pith2026251004157,
  author       = {Pith},
  title        = {Pith review of: GDiffuSE: Diffusion-based speech enhancement with noise model guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFV5KXJY}},
  note         = {Machine review of arXiv:2510.04157}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper introduces a novel speech enhancement (SE) approach based on a denoising diffusion probabilistic model (DDPM), termed Guided diffusion for speech enhancement (GDiffuSE). In contrast to conventional methods that directly map noisy speech to clean speech, our method employs a lightweight helper model to estimate the noise distribution, which is then incorporated into the diffusion denoising process via a guidance mechanism. This design improves robustness by enabling seamless adaptation to unseen noise types and by leveraging large-scale DDPMs originally trained for speech generation in the context of SE. We evaluate our approach on noisy signals obtained by adding noise samples from the BBC sound effects database to LibriSpeech utterances, showing consistent improvements over state-of-the-art baselines under mismatched noise conditions. Examples are available at our project webpage.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 4 linked inside Pith

  1. [1]

    These models perform well under matched conditions but generalize poorly to unseen noise or acoustic environments, often introducing artifacts

    INTRODUCTION Dominant approaches for SE utilize discriminative models that map noisy inputs to clean targets [1]. These models perform well under matched conditions but generalize poorly to unseen noise or acoustic environments, often introducing artifacts. Generative models that learn an explicit prior over clean speech have gained popularity in recent y...

  2. [2]

    , N−1}, andNthe number of samples in the utterance

    PROBLEM FORMULA TION Lety i =x 0,i +w i denote the noisy signal received by a single microphone, wherex 0,i is the clean speech component andw i is the noise component, fori∈ {0, . . . , N−1}, andNthe number of samples in the utterance. Stacking theNsamples into column vectors yieldsx 0 ≜(x 0,i)N−1 i=0 ,w≜(w i)N−1 i=0 ,y≜(y i)N−1 i=0 , leading to the foll...

    Pith/arXiv arXiv 2025

  3. [3]

    PROPOSED METHOD In this section, we derive the proposed SE algorithm. Sec. 3.1 presents the use of DDPM guidance for SE, and Sec. 3.2 describes the training of the noise model that guides the DDPM. The complete process is illustrated in Fig. 1. 3.1. DDPM Guidance for Speech Enhancement DDPM [3] uses a diffusion processs [2] for generative sampling. DDPM g...

  4. [4]

    EXPERIMENTAL STUDY In this section, we provide the implementation details of the pro- posed method, describe the competing method, the datasets used for training and testing, and evaluate the method’s performance. 4.1. Implementation details The noise model architecture is a CNN with 4 causal convolu- tional layers and linear heads forµ t,i andσ t,i, feat...

  5. [5]

    CONCLUSIONS In this work, we introduced GDiffuSE, a lightweight SE method that employs a guidance mechanism to leverage foundation dif- fusion models without retraining the large backbone. By model- ing the noise distribution—an easier task than mapping noisy to clean speech—our approach requires only a short reference noise clip, assuming stable noise st...

  6. [6]

    Supervised speech separation based on deep learning: An overview,

    D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, vol. 26, no. 10, pp. 1702–1726, 2018

    2018

  7. [7]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” inProceedings of the 32nd International Conference on Machine Learning (ICML), 2015

    2015

  8. [8]

    Denoising diffusion probabilis- tic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilis- tic models,”Advances in neural information processing sys- tems, vol. 33, pp. 6840–6851, 2020

    2020

  9. [9]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, P. Abbeel, P. Dhari- wal, and N. B. Chen, “Score-based generative modeling through stochastic differential equations,”International Con- ference on Learning Representations (ICLR), 2021

    2021

  10. [10]

    Dif- fwave: A versatile diffusion model for audio synthesis,

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Dif- fwave: A versatile diffusion model for audio synthesis,” inIn- ternational Conference on Learning Representations (ICLR), 2021

    2021

  11. [11]

    Score-based generative speech enhancement in the complex spectrogram domain,

    C. Welker and W. Kellermann, “Score-based generative speech enhancement in the complex spectrogram domain,” inIEEE In- ternational Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2022, pp. 7317–7321

    2022

  12. [12]

    Universal speech enhancement with score-based diffu- sion models,

    J. Serr `a, J. Pons, P. d. Benito, S. Pascual, and A. Bona- fonte, “Universal speech enhancement with score-based diffu- sion models,”arXiv preprint arXiv:2208.05055, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    A study on speech en- hancement based on diffusion probabilistic model,

    Y .-J. Lu, Y . Tsao, and S. Watanabe, “A study on speech en- hancement based on diffusion probabilistic model,” inAPSIPA Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2021

    2021

  14. [14]

    Spec- Grad: Diffusion-based speech denoising with noisy spectro- gram guidance,

    Y . Koizumi, K. Yatabe, S. Saito, and M. Delcroix, “Spec- Grad: Diffusion-based speech denoising with noisy spectro- gram guidance,” inIEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2022, pp. 8967– 8971

    2022

  15. [15]

    Con- ditional diffusion probabilistic model for end-to-end speech enhancement,

    X. Lu, S. Zhang, K. J. Sim, S. Narayanan, and Z. Li, “Con- ditional diffusion probabilistic model for end-to-end speech enhancement,” inProceedings of the Annual Conference of the International Speech Communication Association (INTER- SPEECH), 2022, pp. 1–5

    2022

  16. [16]

    A connection between score matching and denois- ing autoencoders,

    P. Vincent, “A connection between score matching and denois- ing autoencoders,”Neural Computation, vol. 23, no. 7, pp. 1661–1674, 2011

    2011

  17. [17]

    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

  18. [18]

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

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

    2023

  19. [19]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  20. [20]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,”Advances in neural information processing systems, vol. 35, pp. 26 565– 26 577, 2022

    2022

  21. [21]

    Analysis of classifier-free guid- ance weight schedulers,

    X. Wang, N. Dufour, N. Andreou, M.-P. Cani, V . F. Abrevaya, D. Picard, and V . Kalogeiton, “Analysis of classifier-free guid- ance weight schedulers,”arXiv preprint arXiv:2404.13040, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Applying guidance in a limited interval im- proves sample and distribution quality in diffusion models,

    T. Kynk ¨a¨anniemi, M. Aittala, T. Karras, S. Laine, T. Aila, and J. Lehtinen, “Applying guidance in a limited interval im- proves sample and distribution quality in diffusion models,” Advances in Neural Information Processing Systems, vol. 37, pp. 122 458–122 483, 2024

    2024

  23. [23]

    Wavenet: A generative model for raw au- dio,

    A. van den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, and K. Kavukcuoglu, “Wavenet: A generative model for raw au- dio,”arXiv preprint arXiv:1609.03499, 2016

    Pith/arXiv arXiv 2016

  24. [24]

    Undiff: Unsupervised voice restoration with un- conditional diffusion model,

    A. Iashchenko, P. Andreev, I. Shchekotov, N. Babaev, and D. Vetrov, “Undiff: Unsupervised voice restoration with un- conditional diffusion model,” inProc. Interspeech, 2023

    2023

  25. [25]

    CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” https://doi.org/10.7488/ds/2645, 2019

    doi:10.7488/ds/2645 2019

  26. [26]

    The LJ speech dataset,

    K. Ito and L. Johnson, “The LJ speech dataset,” https://keithito.com/LJ-Speech-Dataset/, 2017

    2017

  27. [27]

    CSR-I (WSJ0) Complete,

    J. S. Garofolo, D. Graff, D. Paul, and D. Pallett, “CSR-I (WSJ0) Complete,” https://catalog.ldc.upenn.edu/LDC93S6A, Linguistic Data Consortium, Philadelphia, 1993

    1993

  28. [28]

    TIMIT acoustic-phonetic continuous speech corpus,

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, N. L. Dahlgren, and V . Zue, “TIMIT acoustic-phonetic continuous speech corpus,” https://catalog.ldc.upenn.edu/LDC93S1, Linguistic Data Consortium, Philadelphia, 1993

    1993

  29. [29]

    The third CHiME speech separation and recognition challenge: Dataset, task and baselines,

    J. Barkeret al., “The third CHiME speech separation and recognition challenge: Dataset, task and baselines,” inProc. IEEE ASRU, 2015, pp. 504–511

    2015

  30. [30]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotovet al., “Librispeech: An ASR corpus based on public domain audio books,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  31. [31]

    BBC sound effects archive,

    “BBC sound effects archive,” https://sound- effects.bbcrewind.co.uk/, 2025

    2025

  32. [32]

    An algorithm for intelligibility prediction of time–frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An algorithm for intelligibility prediction of time–frequency weighted noisy speech,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 7, pp. 2125–2136, 2011

    2011

  33. [33]

    Perceptual evaluation of speech quality (PESQ)—A new method for speech quality assessment of telephone net- works and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hek- stra, “Perceptual evaluation of speech quality (PESQ)—A new method for speech quality assessment of telephone net- works and codecs,” inProc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2001, pp. 749–752

    2001

  34. [34]

    SDR – half-baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR – half-baked or well done?” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019

    2019

  35. [35]

    DNSMOS: A non-intrusive perceptual objective speech qual- ity metric to evaluate noise suppressors,

    C. K. A. Reddy, V . G. Tarunathan, H. Dubey, and et al., “DNSMOS: A non-intrusive perceptual objective speech qual- ity metric to evaluate noise suppressors,” inIEEE Interna- tional Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2021, pp. 6493–6497

    2021

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.