Pith. sign in

REVIEW 2 major objections 5 minor 29 references

On the Application of Diffusion Models for Simultaneous Denoising and Dereverberation

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Diffusion-model cascades for noisy reverberant speech only work if the dominant distortion goes first; a single mixed-objective model is the best default.

desk verdict Solid empirical comparison of cascade vs joint diffusion enhancement, but the low-SNR design rule depends on a synthetic benchmark that isn't validated against real far-field data. read the letter →

arxiv 2508.18833 v1 pith:XR3RVJZA submitted 2025-08-26 eess.AS

Adrian Meise , Tobias Cord-Landwehr , Reinhold Haeb-Umbach This is my paper · ORCID
classification eess.AS
keywords diffusionmodelsspeechenhancementdereverberationdenoisingscore-basedgenerativecascadedmixed-objectivetraininglow-SNRnoisy-reverberant
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

Diffusion models can remove either noise or reverberation well on their own, but real speech often suffers from both at once. This paper asks how to get both out with score-based diffusion enhancers: chain two specialized models, or train one model on both tasks. It finds that cascades are brittle—they only work when the more severe distortion is removed first, and in strongly noisy, strongly reverberant conditions no cascade order saves them. The best option is a single model trained on equal shares of noisy-only, reverberant-only, and noisy-reverberant speech. If true, engineers can skip designing separate denoising and dereverberation stages and use one mixed-objective diffusion model as a universal enhancer.

What carries the argument

The load-bearing mechanism is the score-based diffusion enhancer (SGMSE+) in the complex short-time Fourier transform domain. A forward SDE, dxt = γ(y − xt)dt + g(t)dw, pulls a clean sample toward the observed degraded signal while adding Gaussian noise, and a trained score model solves the reverse SDE to regenerate clean speech from the observation. Because the added Gaussian noise is meant to mask the specific disturbance, the same model can in principle remove noise, reverberation, or both; the paper's experimental machinery is the comparison of four training-data configurations and two cascade orderings that tests this claim.

What would settle it

Run the four trained configurations on real concurrent noise-plus-reverberation recordings with SNR below 10 dB and T60 around 0.7 s, measuring WER and DNSMOS. If the cascade applied in the dominant-distortion order (denoising first when noise dominates, dereverberation first when reverb dominates) performs at least as well as the Joint and Mixed models, the paper's central claim that cascades fail in low-SNR conditions is falsified; if the real-recording results diverge from the synthetic EARS-WHAMR! results, the weakest assumption is exposed.

Watch

Extended reading notes

Core claim

Using the SGMSE+ score-based diffusion architecture, the authors train four enhancers: a denoiser, a dereverberator, a joint model trained only on speech that is both noisy and reverberant, and a mixed-objective model trained on one third of each degradation type. On single-distortion test sets, each dedicated model performs best, but the joint and mixed models stay close. On the real REVERB set (high SNR, reverberation dominant), the dereverberation model alone has the lowest word error rate, the cascade with dereverberation first is solid, and the denoising-first cascade fails. On the new low-SNR synthetic EARS-WHAMR! set, all dedicated and cascaded systems collapse—word error rates rise t

Load-bearing premise

The low-SNR conclusions rest on a new synthetic dataset whose noisy-reverberant mixtures are generated by adding WHAM noise to EARS-Reverb with the same parameterization as the separate datasets; if this simulation does not represent real far-field low-SNR rooms, the cascade failure and single-model advantage may not carry over to real devices.

Editorial extensions

If this is right

  • For real deployments where the dominant distortion is unknown, task-specific cascade pipelines should be replaced by a mixed-objective single model to avoid order-dependent failures.
  • A model trained only on jointly degraded speech is usable when the test condition is known to be noisy and reverberant, but mixed training recovers performance on pure-noise and pure-reverberation conditions.
  • The cascade approach offers no accuracy advantage on joint distortions and costs twice the inference time of a single model.
  • Diffusion models can absorb multiple distortion types into one training distribution without architectural changes, so combining distortions is a data recipe, not a model redesign.
  • On high-SNR reverb-dominant real data, a dedicated dereverberation model plus a robust recognizer can be enough, but perceptual quality remains lower without denoising.

Reading between the lines

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

  • The paper leaves untested whether the same mixed-objective recipe extends to other distortions; adding equal-sized subsets for echo, bandwidth loss, or clipping is a natural cheap extension.
  • The decisive comparison on real low-SNR far-field recordings is still missing; if EARS-WHAMR! is more optimistic than real conditions, the advantage of mixed/joint models could shrink in the very regime where the paper needs it most.
  • The high word error rates even for the best models at negative SNR suggest content preservation, not just naturalness, is the remaining bottleneck; ASR-aware training is a testable extension.
  • The 'dominating distortion first' rule could be made operational by a lightweight SNR/T60 classifier that picks cascade order, which would test the rule without training a joint model.
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

2 major / 5 minor

Summary. The paper investigates how score-based diffusion speech enhancement can be applied to the joint task of denoising and dereverberation. Four model configurations are trained on the SGMSE+ architecture: a denoising model, a dereverberation model, a joint model trained only on noisy-reverberant data, and a mixed-objective model trained on equal shares of noisy, reverberant, and noisy-reverberant data. The systems are evaluated on WSJ-CHiME3, WSJ-Reverb, EARS-WHAM, EARS-Reverb, the real REVERB challenge test set, and a newly synthesized EARS-WHAMR! dataset. The reported results show that cascades of single-task models work only when the cascade is applied in the order of the dominating distortion, and that a single mixed-objective model provides a robust compromise across all conditions, with word error rate as the primary downstream metric. The central claim is that cascade design requires knowing the dominant distortion, and that a mixed-objective diffusion model is the best single-model default.

Significance. The paper addresses a practically important and relatively underexplored problem—joint denoising and dereverberation with diffusion models—and provides the first systematic comparison of cascaded and single-model strategies, including evaluation on the real REVERB recordings. The use of an ASR word error rate in addition to conventional objective metrics is a strength, since it guards against hallucinated but natural-sounding outputs. The large effect sizes observed in the cascade-order experiments (e.g., Table 3a: WER increases from 6.33 to 78.24 when the wrong cascade order is used) make the core qualitative finding credible. However, the low-SNR portion of the design rule rests entirely on a synthetic dataset whose realism is not validated, and the distinction between the joint and mixed-objective models is based on small, unreplicated differences. If the results hold, they would give useful practical guidance to practitioners; in its current form, the evidence supports the findings only under the specific simulated conditions tested.

major comments (2)
  1. [Section 3, Table 1 (EARS-WHAMR!)] The low-SNR conclusions in the abstract and Section 4.2.2 are entirely dependent on the synthetic EARS-WHAMR! dataset. Section 3 states that it was created by adding WHAM noise to EARS-Reverb 'using the same parameterization as for the individual datasets,' but the paper provides no validation that this simulation reproduces real far-field low-SNR noisy-reverberant conditions. No SNR/T60 distribution, no comparison against real recordings in this regime, and no analysis of whether the added noise is reverberated or matches the spectral/temporal structure of real ambient noise are given. Since REVERB, the only real joint test set, has approximately 20 dB SNR with dominant reverberation, the observed cascade failures and joint/mixed advantages (WER 42.6/44.2 vs 80.8/86.7 in Table 1) may be artifacts of the specific synthetic construction. This is load-bearing for the paper's central design
  2. [Section 5, Tables 1 and 3] The conclusion that the mixed-objective model is the 'preferred choice' is not supported by the reported differences. In Table 1, the mixed model has higher WER than the joint model on EARS-WHAMR! (44.2 vs 42.6) and higher WER on REVERB (19.0 vs 17.2), while in Table 3b it has lower WER (9.73 vs 11.69) and in Table 3a nearly identical WER (7.86 vs 7.77). No significance testing or confidence intervals are reported for these pairwise comparisons. The observed differences are small relative to the improvements over the observation and over the cascades, so the claim that mixed objective training is the best single-compromise model is not established. The wording should be softened (e.g., 'comparable to' rather than 'best'), or the authors should provide statistical evidence that the mixed objective is not worse than the joint model across conditions.
minor comments (5)
  1. [Section 3] Please specify the exact parameterization of EARS-WHAMR!: the SNR range, T60 range, whether the noise is passed through the RIR, and the number of utterances. This is also needed to interpret the 'low-SNR' characterization in Section 4.2.2.
  2. [Tables 2 and 3] The arrow notation is rendered inconsistently: 'Den. )Der.' and 'Der. )Den.' should be 'Denoise → Dereverb' and 'Dereverb → Denoise' to match Table 1. As printed, the direction of the cascade is unclear at a glance.
  3. [References] References [3] and [9] duplicate the same ICASSP paper (Lu et al., 'Conditional diffusion probabilistic model for speech enhancement'). Please merge or renumber.
  4. [Section 4.2.1] The phrase 'left column of Table 1' could be clearer as 'left block of Table 1' since the table is organized by dataset columns, not by a single column.
  5. [Section 2.2] The sentence 'The added noise is supposed to mask the characteristics of the disturbances' could benefit from a citing connection to the actual training objective, since the claim is conceptual rather than derived; as written it may overstate the theoretical grounding.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical comparison against external test sets and ASR; central claims are not defined in terms of model outputs.

full rationale

This paper is an empirical evaluation, not a derivation. The central claims—that cascade order should follow the dominating distortion and that a Mixed Objective model is the best single-model compromise—are supported by direct comparisons on external test sets (REVERB, EARS-WHAMR!, WSJ/EARS subsets) using fixed objective metrics and a fixed ASR. There is no fitted parameter renamed as a prediction, no quantity defined in terms of the quantity it claims to explain, and no self-citation chain used to force a conclusion. The EARS-WHAMR! dataset is synthetic, and its construction is a data-generation choice; while the lack of real low-SNR noisy-reverberant validation is a legitimate external-validity concern, it is not circularity under the stated rules. The paper reuses the SGMSE+ architecture and datasets from prior work, but these are code-reproduced/external and are not invoked to forbid alternatives. The only noteworthy weakness—that low-SNR conclusions rest on an unvalidated synthetic set—belongs to correctness risk, not circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical engineering study without a new theoretical construction. The central quantitative choices are the equal one-third mixing ratio for the Mixed Objective model and inherited SGMSE+ hyperparameters; the latter are not free parameters of this work. The transferability of the conclusions rests on the untested assumptions listed above: that the diffusion framework treats all distortions equivalently, that the fixed ASR and objective metrics are reliable, and that the new synthetic EARS-WHAMR! benchmark represents real low-SNR conditions.

free parameters (1)
  • Mixed Objective training data mixing ratio = 1/3 each of noisy, reverberant, and noisy-reverberant subsets
    The recommendation of the Mixed Objective model as the best universal model depends on this single hand-chosen ratio; no ablation over mixing proportions is reported (Section 3).
assumptions (4)
  • domain assumption The forward diffusion process masks the characteristics of the disturbances present in the observed signal, so the same score-model procedure can remove noise, reverberation, or both.
    Section 2.2 states this as the theoretical motivation for joint training; it is a heuristic claim, not proven or separately validated in the paper.
  • domain assumption SGMSE+ with the same parametrization as earlier work [10,18] is an adequate base system for comparing all configurations.
    All conclusions are conditional on this single architecture and training recipe; no other diffusion system or discriminative baseline is tested.
  • domain assumption DNSMOS, SRMR, ESTOI, and the fixed NeMo QuartzNet WER jointly capture speech quality, dereverberation, intelligibility, and hallucinations.
    Section 3 argues for this combined assessment; however, the ASR is not robust to distortions and may confound enhancement artifacts with recognition errors.
  • domain assumption EARS-WHAMR! synthetic data is representative of real low-SNR noisy reverberant conditions.
    Section 3 describes generating it by adding WHAM noise to EARS-Reverb, but no validation against real low-SNR far-field recordings is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Application of Diffusion Models for Simultaneous Denoising and Dereverberation." pith.science (2026). https://pith.science/paper/XR3RVJZA

@misc{pith2026250818833,
  author       = {Pith},
  title        = {Pith review of: On the Application of Diffusion Models for Simultaneous Denoising and Dereverberation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XR3RVJZA}},
  note         = {Machine review of arXiv:2508.18833}
}
read the original abstract

Diffusion models have been shown to achieve natural-sounding enhancement of speech degraded by noise or reverberation. However, their simultaneous denoising and dereverberation capability has so far not been studied much, although this is arguably the most common scenario in a practical application. In this work, we investigate different approaches to enhance noisy and/or reverberant speech. We examine the cascaded application of models, each trained on only one of the distortions, and compare it with a single model, trained either solely on data that is both noisy and reverberated, or trained on data comprising subsets of purely noisy, of purely reverberated, and of noisy reverberant speech. Tests are performed both on artificially generated and real recordings of noisy and/or reverberant data. The results show that, when using the cascade of models, satisfactory results are only achieved if they are applied in the order of the dominating distortion. If only a single model is desired that can operate on all distortion scenarios, the best compromise appears to be a model trained on the aforementioned three subsets of degraded speech data.

Figures

Figures reproduced from arXiv: 2508.18833 by the authors.

Figure 1
Figure 1. Overview of the possible approaches for com [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

29 extracted references · 27 canonical work pages

  1. [1]

    Deep neural network techniques for monaural speech enhancement and separation: state of the art anal- ysis,

    P. Ochieng, “Deep neural network techniques for monaural speech enhancement and separation: state of the art anal- ysis,” Artificial Intelligence Review , vol. 56, no. Suppl 3, pp. 3651–3703, 2023

    work page 2023

  2. [2]

    Microphone array signal processing and deep learning for speech enhancement: Combining model- based and data-driven approaches to parameter estimation and filtering,

    R. Haeb-Umbach, T. Nakatani, M. Delcroix, C. Boeddeker, and T. Ochiai, “Microphone array signal processing and deep learning for speech enhancement: Combining model- based and data-driven approaches to parameter estimation and filtering,” IEEE Signal Processing Magazine , vol. 41, no. 6, pp. 12–23, 2024

    work page 2024

  3. [3]

    Conditional diffusion probabilistic model for speech enhancement,

    Y .-J. Lu, Z. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional diffusion probabilistic model for speech enhancement,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 7402–7406, 2022

    work page 2022

  4. [4]

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

    S. Welker, J. Richter, and T. Gerkmann, “Speech enhance- ment with score-based generative models in the complex stft domain,” in Proceedings ISCA Interspeech, pp. 2928– 2932, 2022

    work page 2022

  5. [5]

    Diffusion-based speech enhancement in matched and mismatched conditions using a heun-based sampler,

    P. Gonzalez, Z.-H. Tan, J. Østergaard, J. Jensen, T. S. Al- strøm, and T. May, “Diffusion-based speech enhancement in matched and mismatched conditions using a heun-based sampler,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 10431–10435, 2023

    work page 2023

  6. [6]

    Schrödinger bridge for generative speech enhancement,

    A. Juki ´c, R. Korostik, J. Balam, and B. Ginsburg, “Schrödinger bridge for generative speech enhancement,” in Proceedings ISCA Interspeech, pp. 1175–1179, 2024

    work page 2024

  7. [7]

    Analysing diffusion-based generative approaches versus discriminative approaches for speech restoration,

    J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Analysing diffusion-based generative approaches versus discriminative approaches for speech restoration,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, 2023

    work page 2023

  8. [8]

    Unsupervised Blind Joint Dereverberation and Room Acoustics Estimation with Diffusion Models

    J.-M. Lemercier, E. Moliner, S. Welker, V . Välimäki, and T. Gerkmann, “Unsupervised blind joint dereverberation and room acoustics estimation with diffusion models,” arXiv preprint arXiv:2408.07472, 2024

    work page Pith review arXiv 2024

Show all 29 references
  1. [9]

    Conditional diffusion probabilistic model for speech enhancement,

    Y .-J. Lu, Z.-Q. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional diffusion probabilistic model for speech enhancement,” in IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pp. 7402–7406, 2022

    2022

  2. [10]

    Speech enhancement and dereverberation with diffusion-based generative models,

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

    2023

  3. [11]

    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 Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2724–2737, 2023

    2023

  4. [12]

    Uni- versal score-based speech enhancement with high content preservation,

    R. Scheibler, Y . Fujita, Y . Shirahata, and T. Komatsu, “Uni- versal score-based speech enhancement with high content preservation,” in Proceedings ISCA Interspeech, pp. 1165– 1169, 2024

    2024

  5. [13]

    Diffusion posterior sampling for informed single-channel dereverber- ation,

    J.-M. Lemercier, S. Welker, and T. Gerkmann, “Diffusion posterior sampling for informed single-channel dereverber- ation,” in Proc. IEEE WASPAA, pp. 1–5, 2023

    2023

  6. [14]

    Interspeech 2025 urgent speech enhance- ment challenge,

    K. Saijo, W. Zhang, S. Cornell, R. Scheibler, C. Li, Z. Ni, A. Kumar, M. Sach, Y . Fu, W. Wang, T. Fingscheidt, and S. Watanabe, “Interspeech 2025 urgent speech enhance- ment challenge,” 2025

    2025

  7. [15]

    Diffusion model-based mimo speech denoising and dereverberation,

    R. Kimura, T. Nakatani, N. Kamo, D. Marc, S. Araki, T. Ueda, and S. Makino, “Diffusion model-based mimo speech denoising and dereverberation,” in International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pp. 455–459, 2024

    2024

  8. [16]

    Multi- stream diffusion model for probabilistic integration of model-based and data-driven speech enhancement,

    T. Nakatani, N. Kamo, M. Delcroix, and S. Araki, “Multi- stream diffusion model for probabilistic integration of model-based and data-driven speech enhancement,” in Proc. IEEE IWAENC, pp. 65–69, 2024

    2024

  9. [17]

    usee: Unified speech enhancement and editing with conditional diffusion models,

    M. Yang, C. Zhang, Y . Xu, Z. Xu, H. Wang, B. Raj, and D. Yu, “usee: Unified speech enhancement and editing with conditional diffusion models,” IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pp. 7125–7129, 2023

    2023

  10. [18]

    Ears: An anechoic full- band speech dataset benchmarked for speech enhancement and dereverberation,

    J. Richter, Y .-C. Wu, S. Krenn, S. Welker, B. Lay, S. Watan- abe, A. Richard, and T. Gerkmann, “Ears: An anechoic full- band speech dataset benchmarked for speech enhancement and dereverberation,” in Proceedings ISCA Interspeech , pp. 4873–4877, 2024

    2024

  11. [19]

    The REVERB challenge: A common evaluation framework for derever- beration and recognition of reverberant speech,

    K. Kinoshita, M. Delcroix, T. Yoshioka, T. Nakatani, E. Ha- bets, R. Haeb-Umbach, V . Leutnant, A. Sehr, W. Keller- mann, R. Maas, S. Gannot, and B. Raj, “The REVERB challenge: A common evaluation framework for derever- beration and recognition of reverberant speech,” in Proc....

    2013

  12. [20]

    A connection between score matching and de- noising autoencoders,

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

    2011

  13. [21]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in International Conference on Learning Representations, 2021

    2021

  14. [22]

    Speech En- hancement and Dereverberation with Diffusion-based Gen- erative Models

    Signal Processing (SP), Universität Hamburg, “Speech En- hancement and Dereverberation with Diffusion-based Gen- erative Models.” https://github.com/sp-uhh/ sgmse/, 2025

    2025

  15. [23]

    CSR-I (WSJ0) Complete LDC93S6A,

    J. Garofolo, D. Graff, D. Paul, and D. Pallett, “CSR-I (WSJ0) Complete LDC93S6A,” Web Download, Philadel- phia: Linguistic Data Consortium, 1993

    1993

  16. [24]

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

    C. K. A. Reddy, V . Gopal, and R. Cutler, “Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pp. 6493–6497, 2021

    2021

  17. [25]

    A non-intrusive quality and intelligibility measure of reverberant and dere- verberated speech,

    T. H. Falk, C. Zheng, and W.-Y . Chan, “A non-intrusive quality and intelligibility measure of reverberant and dere- verberated speech,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, no. 7, pp. 1766–1774, 2010

    2010

  18. [26]

    An algorithm for predicting the intelligibility of speech masked by modulated noise maskers,

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

    2009

  19. [27]

    SDR - Half-baked or Well Done?,

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

    2019

  20. [28]

    Sms-wsj: Database, performance measures, and baseline recipe for multi-channel source separation and recognition,

    L. Drude, J. Heitkaemper, C. Boeddeker, and R. Haeb- Umbach, “Sms-wsj: Database, performance measures, and baseline recipe for multi-channel source separation and recognition,” arXiv preprint arXiv:1910.13934, 2019

    1910 arXiv

  21. [29]

    Nemo: a toolkit for building AI applications using neural modules,

    O. Kuchaiev, J. Li, H. Nguyen, O. Hrinchuk, R. Leary, B. Ginsburg, S. Kriman, S. Beliaev, V . Lavrukhin, J. Cook, P. Castonguay, M. Popova, J. Huang, and J. M. Cohen, “Nemo: a toolkit for building AI applications using neural modules,” CoRR, vol. abs/1909.09577, 2019

    1909 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.