Pith. sign in

REVIEW 3 major objections 6 minor 76 references

VINP: Variational Bayesian Inference with Neural Speech Prior for Joint ASR-Effective Speech Dereverberation and Blind RIR Identification

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A dereverberation DNN's output can be reused as a Gaussian prior inside a variational Bayesian loop, yielding state-of-the-art speech cleanup and room impulse response estimates from one recording.

desk verdict Strong ASR-focused dereverberation; the blind RT60 SOTA claim is undercut by a self-referential evaluation and finite-CTF truncation. read the letter →

arxiv 2502.07205 v3 pith:ILST4ZTV submitted 2025-02-11 eess.AS cs.AIcs.LG

Pengyu Wang , Ying Fang , Xiaofei Li This is my paper · ORCID
classification eess.AScs.AIcs.LG
keywords speechdereverberationblindRIRidentificationvariationalBayesianinferenceconvolutivetransferfunctionneuralpriorASR-effectivefront-endRT60estimationDRR
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

Dereverberated speech and the room that produced it are usually estimated by separate tools, but this paper argues that both can be solved together in one probabilistic framework. The key move is to take an off-the-shelf discriminative speech-enhancement network, retrain only its loss so that its output is a Gaussian prior over the clean speech magnitude, and then run variational Bayesian expectation-maximization on a convolutive-transfer-function model of the reverberant recording. The loop produces a maximum-a-posteriori clean spectrum and a maximum-likelihood estimate of the room filter, which a pseudo sine-sweep measurement converts into a room impulse response waveform. On the REVERB test set the authors report top listening scores and word error rates, and on their SimACE test set the best blind RT60 accuracy, with computational cost linear in utterance length.

What carries the argument

The load-bearing object is the convolutional transfer function (CTF) approximation, which models reverberation in each STFT band as a short filter across frames, $X(f,t)\approx\sum_{l=0}^{L-1}H_l(f)S(f,t-l)+W(f,t)$, rather than as a full cross-band filter bank. On top of this, the machinery is variational expectation-maximization: a mean-field Gaussian posterior for the clean spectrum with closed-form E-step updates for mean and precision, M-step updates for noise precision and CTF coefficients, and an exponential moving average for stable convergence. The DNN prior is trained with a KL-divergence (Itakura-Saito-type) loss so that its magnitude output defines $\alpha(f,t)=1/|\hat S_N(f,t)|^2$, and it stays fixed during VEM, giving single-forward-pass DNN inference. A pseudo measurement generated by convolving the CTF filter with a logarithmic sine sweep, then inverse filtering, converts the CTF estimate into an RIR waveform.

What would settle it

Synthesize reverberant speech from a measured RIR with an RT60 well above 1.5 s and with known late-tail energy, run VINP, and compare the estimated RIR's RT60 and DRR with ground truth; if the error systematically grows with the energy that lies beyond the 30-frame CTF window, or if injecting cross-band filter coefficients causes a large bias, the finite band-to-band CTF premise is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a probabilistic signal model and a discriminative DNN are not competing approaches: a DNN's estimate of the clean speech magnitude can be interpreted as the precision of a complex-Gaussian prior, and then the observation likelihood of a CTF model corrects that prior through closed-form variational updates. The MAP estimate of the anechoic spectrum and the ML estimate of the CTF filter are obtained analytically per frequency band, so no iterative DNN inference or MCMC sampling is needed; the DNN runs once and the VEM iterations are parallel across time-frequency bins. Reconstructing the RIR waveform is done by feeding the estimated CTF filter with a logarithmic sine sweep and applying the sweep's inverse filter. The paper reports that this design makes ASR-ineffective DNNs ASR-effective and makes ASR-effective DNNs better, and it reports SOTA MOS/WER on REVERB and SOTA RT60 estimation on SimACE.

Load-bearing premise

The entire estimate rests on the assumption that, in each frequency band, reverberation is fully described by a short chain of previous STFT frames (about 0.24 seconds with $L=30$) with no cross-band mixing; if a room's echo tail is longer or carries significant cross-band energy, both the cleaned speech and the room impulse response estimate will be biased.

Editorial extensions

If this is right

  • A front-end that respects an explicit signal model can make a speech-enhancement DNN safe for downstream ASR even when that DNN's raw output hurts ASR.
  • The same recordings that are dereverberated also yield an estimated room impulse response, so the method can supply room acoustic parameters (RT60, DRR) without a separate calibration step.
  • The computational cost of the Bayesian loop scales linearly with utterance duration, in contrast to cubic scaling in the prior VAE-EM approach, and the loop parallelizes across time-frequency bins.
  • Because the prior is fixed during inference, any improved discriminative dereverberation network can be plugged in as the backbone without changing the Bayesian update rules.
  • The method's outputs remain consistent with the CTF signal model, which is the property the authors credit for reduced DNN artifacts and improved ASR.

Reading between the lines

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

  • The same 'DNN output as prior, then model-based correction' recipe could transfer to other inverse problems, such as speech denoising or source separation, whenever a cheap discriminative estimator exists and a probabilistic observation model is available; the paper only demonstrates the idea for dereverberation.
  • A natural extension would be to let the speech prior be updated during the VEM loop instead of fixing it, trading the single-forward-pass property for a potentially stronger prior; the paper does not explore this direction.
  • The finite CTF length of $L=30$ frames, noted by the authors as a limitation, suggests that very reverberant rooms with RT60 above roughly 1.5 s may require a longer filter or an explicit late-reverberation tail, an extension the paper leaves implicit.
  • The pseudo sine-sweep bandwidth (62.5 Hz to 8 kHz) and its inverse filter set the usable frequency range of the estimated RIR; varying the sweep bandwidth would be a direct test of how much of the room response VINP can recover.
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

3 major / 6 minor

Summary. The paper proposes VINP, a variational Bayesian inference framework that combines a discriminative DNN-based speech prior with a CTF-based probabilistic signal model for joint speech dereverberation and blind RIR identification. The VEM derivation is clean, and the dereverberation experiments on REVERB support the claim that the method is ASR-effective, including a subjective listening test and WER evaluations with multiple Whisper models. The blind RIR identification experiments on SimACE report low RT60/DRR errors, but the evaluation is weakened by the finite CTF truncation and by the fact that the reference acoustic parameters are computed with the same heuristic used for the estimates.

Significance. If the dereverberation results hold, VINP is a valuable contribution: it shows that a supervised DNN output can be used as a Bayesian prior inside a variational loop, improving ASR robustness while retaining a linear-complexity, parallelizable inference procedure. The paper ships code and audio samples, reports detailed training settings, and includes a broad set of speech-quality and ASR metrics. The blind RIR identification half of the central claim, however, is not yet supported at the same level of certainty: the finite CTF model and the self-referential RT60/DRR evaluation create a real risk that the reported accuracy reflects truncated-model consistency rather than identification of the full RIR. The contribution would be substantially strengthened by either fixing this evaluation or re-scoping the RIR claim.

major comments (3)
  1. [Section IV-E2 and Section IV-B4] The CTF filter length L=30, together with the stated STFT settings (512-sample Hann window, 75% overlap), covers approximately 0.264 s of the room impulse response, while SimACE contains RIRs with RT60 up to 1.22 s. The paper argues in Section IV-E2 that the late-tail energy is 'relatively negligible', but this is not quantitatively supported: for an RT60 of 1.22 s, the energy beyond 0.264 s is roughly 5% of the total and the EDC is only about 13 dB below the direct sound. This is not negligible for RT60 estimation and can bias both the MAP speech estimate and the RIR estimate. Please provide a quantitative analysis of truncation error (e.g., EDC of the truncated versus full RIR), extend the CTF length, or explicitly re-scope the claim to short-RIR identification.
  2. [Section IV-D2 and Table III] The RT60 and DRR reference values in the SimACE evaluation are computed with the same heuristic equations (38)-(39) that are applied to the estimates. The fitting range in Eq. (38) is a short 5 dB segment immediately after the direct path, so a CTF-truncated RIR can produce a high-correlation straight early-decay segment even if the late tail is completely missing. The low MAE in Table III may therefore reflect self-consistency between the truncated model and the heuristic reference rather than successful identification of the full RIR. Please report errors against ground-truth RT60/DRR values computed from the full measured RIRs, or supplement the table with EDC or full-RIR waveform errors.
  3. [Table III] No confidence intervals or statistical significance tests are reported for the RT60 and DRR MAE/RMSE values. Given the limited size of SimACE and the visible trade-off between the two VINP variants (better RT60 for VINP-TCN+SA+S, better DRR for VINP-oSpatialNet), the differences among methods may not be meaningful. Please provide per-utterance error distributions or confidence intervals for all RIR-identification metrics.
minor comments (6)
  1. [Table VI] The header 'MAE of RT60 estiamtion' contains a typo: 'estiamtion' should be 'estimation'.
  2. [Figure 4] The y-axis label 'MACs(sIteration)-1 (G)' is unclear; please rewrite it to indicate the unit, e.g., 'MACs per second per iteration (G/s/it)'.
  3. [Section III-A] The phrase 'arbitrary discriminative dereverberation DNN' is stronger than what is implemented: the backbone is trained with a modified KL loss and, in the oSpatialNet case, with a modified architecture. Suggest wording such as 'any dereverberation DNN adapted with the proposed prior-estimation loss'.
  4. [Table II] The claim that VINP-oSpatialNet achieves SOTA in ASR should be qualified: on SimData with the Whisper 'medium' model it ties with oSpatialNet* at 4.2% WER, so the advantage is not uniform across all conditions.
  5. [Section IV-E1] The subjective MOS test uses only 10 input utterances. Please report confidence intervals or a statistical test for the MOS differences to substantiate the SOTA MOS claim.
  6. [Section III-B and Conclusion] The statement that VINP 'avoids the direct utilization of DNN output' is misleading because the DNN output directly sets the prior precision α(f,t) in Eq. (13). Consider phrasing it as 'does not use the DNN output as the final enhanced spectrum'.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; one benchmark-level circularity in RT60/DRR reference construction, while the core VEM derivation is self-contained.

  1. other [Section IV-D2 (Evaluation Metrics for Blind RIR Identification), with Eqs. (38)-(39) and Section IV-C2]
    "Notice that the reference RT60s and DRRs are calculated using Eq. (38) and Eq. (39) to ensure fair comparison."

    The blind RIR identification results are scored against reference RT60 and DRR values that are themselves computed with the same heuristic equations used to score the estimates: Schroeder EDC linear fitting over a short 5 dB segment for RT60 (Eq. (38)) and a fixed direct-path window for DRR (Eq. (39)). Thus the reported MAE/RMSE measure consistency of the heuristic applied to the estimated and true RIRs, not agreement with an independently measured physical parameter. The ground truth is defined by the same estimator family as the output, which is a benchmark-level circularity. This does not enter the VEM derivation or the pseudo-measurement RIR reconstruction, so the central dereverberation and CTF-estimation claims remain independent.

full rationale

The core derivation is a standard variational Bayesian EM procedure: the DNN supplies a Gaussian speech prior via alpha_N = 1/|S_hat_N|^2 (Eq. (13)), the E-step computes a closed-form Gaussian posterior (Eq. (21)) using the observation likelihood, and the M-step estimates the CTF filter by maximum likelihood (Eq. (25)). No fitted parameter is renamed as a prediction, and the MAP speech estimate explicitly combines prior and observation rather than simply returning the DNN output. The CTF-to-RIR reconstruction is a deterministic pseudo-measurement transform (Eqs. (35)-(36)) and is not fitted to the RT60/DRR references. The only self-referential element is the RIR evaluation protocol: the reference RT60s and DRRs are calculated with the same heuristic Eqs. (38)-(39) used to score the estimates, which can inflate agreement but does not circularly determine the estimates themselves. The paper also explicitly acknowledges the CTF finite-length limitation in Section IV-E2 ('a CTF filter with a finite length can only model a RIR with a finite length'), a correctness caveat rather than a derivation circularity. Overall, the central VBI derivation is self-contained; the mild evaluation-level circularity does not affect the speech dereverberation or ASR claims.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central derivation rests on the CTF approximation, Gaussian independence assumptions, and the DNN-as-prior replacement. The manually chosen hyperparameters are not exotic, but several (L, lambda, iteration count, RT60 fitting constants) influence the reported results. No new physical entities are introduced. The main epistemic load is carried by the assumption that the DNN prior is a faithful stand-in for the true speech prior and by the pseudo measurement reconstruction of the RIR from a finite CTF.

free parameters (5)
  • CTF filter length L = 30 STFT frames (about 0.24 s)
    Chosen empirically; it truncates the modeled RIR and directly bounds the late-reverberation information available for dereverberation and RIR identification.
  • EMA smoothing factor lambda = 0.7
    Hand-set in Eq. (22) to stabilize VEM iteration; the posterior update is a convex combination of previous and current estimates, and performance depends on this balance.
  • VEM iteration count = 100 for dereverberation, 300 for RIR identification
    Empirically set; Fig. 5 and Fig. 8 show that likelihood and metrics continue changing with iteration count, so this is a tuned hyperparameter.
  • Lowest ignored frequency bands = 3 bands (254 bands processed)
    Set to zero to avoid bins below the 85 Hz fundamental where SNR is very low; this choice affects which information enters the VBI updates.
  • RT60/DRR fitting heuristic constants = 5 dB drop, 50 ms start window, 2.5 ms direct-path spread
    These constants define the evaluation rule in Eq. (38) and Eq. (39); both the reference and the estimates use the same rule, which makes the metric sensitive to the chosen heuristic.
assumptions (5)
  • domain assumption The reverberant observation is modeled by a band-to-band CTF approximation, X(f,t) = H(f)S(f,t) + W(f,t), ignoring cross-band filters.
    Eq. (2); the approximation is standard and cited to [39] and [43], but its error is not separately quantified in VINP.
  • domain assumption Speech and noise are zero-mean complex Gaussian with precisions alpha and delta, independent across all T-F bins.
    Eq. (3)-(5); Assumptions 1-2. This factorization enables the mean-field posterior in Eq. (15), but real speech has strong time-frequency correlations.
  • ad hoc to paper The discriminative DNN output approximates the oracle anechoic speech prior, so alpha(f,t) = 1/|S_hat_N(f,t)|^2 is used as the precision in the VBI loop.
    Section III-A, Eq. (11)-(13); any DNN error becomes prior error, and the prior is not updated during iterations.
  • domain assumption CTF applied to a known excitation signal in the STFT domain can be inverted to recover the RIR waveform via inverse filtering.
    Section III-C, Algorithm 2 and Eq. (31)-(36); the paper provides a pseudo measurement construction rather than a formal proof.
  • domain assumption The minimum power over frames initializes the noise precision.
    Eq. (30), based on minimum statistics noise estimation [47]; assumes the noise floor appears in at least one frame.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VINP: Variational Bayesian Inference with Neural Speech Prior for Joint ASR-Effective Speech Dereverberation and Blind RIR Identification." pith.science (2026). https://pith.science/paper/ILST4ZTV

@misc{pith2026250207205,
  author       = {Pith},
  title        = {Pith review of: VINP: Variational Bayesian Inference with Neural Speech Prior for Joint ASR-Effective Speech Dereverberation and Blind RIR Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILST4ZTV}},
  note         = {Machine review of arXiv:2502.07205}
}
read the original abstract

Reverberant speech, denoting the speech signal degraded by reverberation, contains crucial knowledge of both anechoic source speech and room impulse response (RIR). This work proposes a variational Bayesian inference (VBI) framework with neural speech prior (VINP) for joint speech dereverberation and blind RIR identification. In VINP, a probabilistic signal model is constructed in the time-frequency (T-F) domain based on convolution transfer function (CTF) approximation. For the first time, we propose using an arbitrary discriminative dereverberation deep neural network (DNN) to estimate the prior distribution of anechoic speech within a probabilistic model. By integrating both reverberant speech and the anechoic speech prior, VINP yields the maximum a posteriori (MAP) and maximum likelihood (ML) estimations of the anechoic speech spectrum and CTF filter, respectively. After simple transformations, the waveforms of anechoic speech and RIR are estimated. VINP is effective for automatic speech recognition (ASR) systems, which sets it apart from most deep learning (DL)-based single-channel dereverberation approaches. Experiments on single-channel speech dereverberation demonstrate that VINP attains state-of-the-art (SOTA) performance in mean opinion score (MOS) and word error rate (WER). For blind RIR identification, experiments demonstrate that VINP achieves SOTA performance in estimating reverberation time at 60 dB (RT60) and advanced performance in direct-to-reverberation ratio (DRR) estimation. Codes and audio samples are available online.

Figures

Figures reproduced from arXiv: 2502.07205 by the authors.

Figure 1
Figure 1. Probabilistic graphical model in VINP. B. Task Description Our goal is to jointly achieve speech dereverberation and blind RIR identification by solving all hidden variables and parameters in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of VINP. A. Estimation of Anechoic Speech Prior The direct-path speech signal, which is a scaled and delayed version of the anechoic source speech signal, is free from noise and reverberation. To avoid estimating the arbitrary direct-path propagation delay and attenuation, instead of the actual source speech, we setup the direct-path speech as the source speech and our target signal, which is still denoted … view at source ↗
Figure 3
Figure 3. DNN architectures in VINP. 2) DNN Architecture: VINP is able to employ any discrim￾inative dereverberation DNNs to estimate the prior distribu￾tion of anechoic speech. Two representative dereverberation backbones TCN+SA+S [18] and oSpatialNet-Mamba [60] are selected. Among those backbones, TCN+SA+S includes a self￾attention module to produce dynamic representations given input features, a temporal convolutional netw… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: MACs per second and per iteration versus speech [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Logarithmic likelihood and dereverberation metrics as a function of VEM iterations on REVERB SimData. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: MOS of subject listening test. are significant, except for VINP-oSpatialNet and oSpatialNet* on the Whisper ’medium’ model. When comparing VINP w/o VEM with the original backbones, it becomes evident that the DNNs adapted to speech prior estimation in the proposed VINP…
Figure 8
Figure 8. Figure 8: MAE of RT60 and DRR estimation as a function of [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 7
Figure 7. Figure 7: Examples of magnitude spectra (RT60=0.7 s). [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 59 canonical work pages

  1. [1]

    Towards improved room impulse response estimation for speech recognition,

    A. Ratnarajah, I. Ananthabhotla, V . K. Ithapu, P. Hoffmann, D. Manocha, and P. Calamia, “Towards improved room impulse response estimation for speech recognition,” inICASSP 2023-2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    work page 2023

  2. [2]

    Vincent, T

    E. Vincent, T. Virtanen, and S. Gannot,Audio source separation and speech enhancement. John Wiley & Sons, 2018

    work page 2018

  3. [3]

    High-order diffraction and dif- fuse reflections for interactive sound propagation in large environments,

    C. Schissler, R. Mehra, and D. Manocha, “High-order diffraction and dif- fuse reflections for interactive sound propagation in large environments,” ACM Transactions on Graphics (TOG), vol. 33, no. 4, pp. 1–12, 2014

    work page 2014

  4. [4]

    Speech dereverberation based on variance-normalized delayed linear prediction,

    T. Nakatani, T. Yoshioka, K. Kinoshita, M. Miyoshi, and B.-H. Juang, “Speech dereverberation based on variance-normalized delayed linear prediction,”IEEE Transactions on Audio, Speech, and Language Pro- cessing, vol. 18, no. 7, pp. 1717–1731, 2010

    2010

  5. [5]

    Blind speech dereverberation with multi-channel linear prediction based on short time fourier transform representation,

    ——, “Blind speech dereverberation with multi-channel linear prediction based on short time fourier transform representation,” in2008 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2008, pp. 85–88

    work page 2008

  6. [6]

    Suppression of late reverberation effect on speech signal using long-term multiple- step linear prediction,

    K. Kinoshita, M. Delcroix, T. Nakatani, and M. Miyoshi, “Suppression of late reverberation effect on speech signal using long-term multiple- step linear prediction,”IEEE transactions on audio, speech, and lan- guage processing, vol. 17, no. 4, pp. 534–545, 2009

    work page 2009

  7. [7]

    Generalization of multi-channel linear prediction methods for blind mimo impulse response shortening,

    T. Yoshioka and T. Nakatani, “Generalization of multi-channel linear prediction methods for blind mimo impulse response shortening,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, no. 10, pp. 2707–2720, 2012

    work page 2012

  8. [8]

    Frequency-domain single- channel inverse filtering for speech dereverberation: Theory and prac- tice,

    I. Kodrasi, T. Gerkmann, and S. Doclo, “Frequency-domain single- channel inverse filtering for speech dereverberation: Theory and prac- tice,” in2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2014, pp. 5177–5181

    work page 2014

Show all 76 references
  1. [9]

    Vari- ational bayesian inference for multichannel dereverberation and noise reduction,

    D. Schmid, G. Enzner, S. Malik, D. Kolossa, and R. Martin, “Vari- ational bayesian inference for multichannel dereverberation and noise reduction,”IEEE/ACM transactions on audio, speech, and language processing, vol. 22, no. 8, pp. 1320–1335, 2014

    2014

  2. [10]

    Speech dereverberation using non- negative convolutive transfer function and spectro-temporal modeling,

    N. Mohammadiha and S. Doclo, “Speech dereverberation using non- negative convolutive transfer function and spectro-temporal modeling,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 24, no. 2, pp. 276–289, 2015

    2015

  3. [11]

    A bayesian hierarchical model for speech enhancement,

    Y . Laufer and S. Gannot, “A bayesian hierarchical model for speech enhancement,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 46–50

    2018

  4. [12]

    A bayesian hierarchical model for speech enhancement with time- varying audio channel,

    ——, “A bayesian hierarchical model for speech enhancement with time- varying audio channel,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 1, pp. 225–239, 2019

    2019

  5. [13]

    Off-grid direction of arrival estimation using sparse bayesian inference,

    Z. Yang, L. Xie, and C. Zhang, “Off-grid direction of arrival estimation using sparse bayesian inference,”IEEE transactions on signal process- ing, vol. 61, no. 1, pp. 38–43, 2012

    2012

  6. [14]

    An off-grid wideband doa estimation method with the variational bayes expectation-maximization frame- work,

    P. Wang, H. Yang, and Z. Ye, “An off-grid wideband doa estimation method with the variational bayes expectation-maximization frame- work,”Signal Processing, vol. 193, p. 108423, 2022

    2022

  7. [15]

    Joint estimation of direction-of- arrival and distance for arrays with directional sensors based on sparse bayesian learning,

    P. Wang, F. Xiong, Z. Ye, and J. Feng, “Joint estimation of direction-of- arrival and distance for arrays with directional sensors based on sparse bayesian learning,” inINTERSPEECH, 2022

    2022

  8. [16]

    Learning spectral mapping for speech dereverberation and denoising,

    K. Han, Y . Wang, D. Wang, W. S. Woods, I. Merks, and T. Zhang, “Learning spectral mapping for speech dereverberation and denoising,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, no. 6, pp. 982–992, 2015

    2015

  9. [17]

    Real-time single-channel dereverberation and separation with time-domain audio separation network

    Y . Luo and N. Mesgarani, “Real-time single-channel dereverberation and separation with time-domain audio separation network.” inInterspeech, 2018, pp. 342–346

    2018

  10. [18]

    Monaural speech derever- beration using temporal convolutional networks with self attention,

    Y . Zhao, D. Wang, B. Xu, and T. Zhang, “Monaural speech derever- beration using temporal convolutional networks with self attention,” IEEE/ACM transactions on audio, speech, and language processing, vol. 28, pp. 1598–1607, 2020

    2020

  11. [19]

    Fullsubnet: A full-band and sub- band fusion model for real-time single-channel speech enhancement,

    X. Hao, X. Su, R. Horaud, and X. Li, “Fullsubnet: A full-band and sub- band fusion model for real-time single-channel speech enhancement,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6633–6637

    2021

  12. [20]

    Speech dereverberation with a reverbera- tion time shortening target,

    R. Zhou, W. Zhu, and X. Li, “Speech dereverberation with a reverbera- tion time shortening target,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  13. [21]

    Spectro-temporal subnet for real-time monaural speech denoising and dereverberation

    F. Xiong, W. Chen, P. Wang, X. Li, and J. Feng, “Spectro-temporal subnet for real-time monaural speech denoising and dereverberation.” in Interspeech, 2022, pp. 931–935. 13

    2022

  14. [22]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in ICLR, 2014

    2014

  15. [23]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

    2020

  16. [24]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  17. [25]

    Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement,

    S.-W. Fu, C.-F. Liao, Y . Tsao, and S.-D. Lin, “Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement,” inInternational Conference on Machine Learn- ing. PmLR, 2019, pp. 2031–2041

    2019

  18. [26]

    Cmgan: Conformer-based metric- gan for monaural speech enhancement,

    S. Abdulatif, R. Cao, and B. Yang, “Cmgan: Conformer-based metric- gan for monaural speech enhancement,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  19. [27]

    Speech enhancement and dereverberation with diffusion-based genera- tive models,

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

    2023

  20. [28]

    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, 2023

    2023

  21. [29]

    How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,

    K. Iwamoto, T. Ochiai, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,” inInterspeech 2022, 2022, pp. 5418–5422

    2022

  22. [30]

    How does end-to-end speech recognition training impact speech enhancement artifacts?

    ——, “How does end-to-end speech recognition training impact speech enhancement artifacts?” inICASSP 2024-2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 031–11 035

    2024

  23. [31]

    A non-negative approach to semi- supervised separation of speech from noise with the use of temporal dynamics,

    G. J. Mysore and P. Smaragdis, “A non-negative approach to semi- supervised separation of speech from noise with the use of temporal dynamics,” in2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2011, pp. 17–20

    2011

  24. [32]

    Statistical speech enhancement based on probabilistic integration of variational autoencoder and non-negative matrix factorization,

    Y . Bando, M. Mimura, K. Itoyama, K. Yoshii, and T. Kawahara, “Statistical speech enhancement based on probabilistic integration of variational autoencoder and non-negative matrix factorization,” in2018 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing...

    2018

  25. [33]

    Unsuper- vised speech enhancement using dynamical variational autoencoders,

    X. Bie, S. Leglaive, X. Alameda-Pineda, and L. Girin, “Unsuper- vised speech enhancement using dynamical variational autoencoders,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2993–3007, 2022

    2022

  26. [34]

    Speech dereverberation using variational autoencoders,

    D. Baby and H. Bourlard, “Speech dereverberation using variational autoencoders,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 5784–5788

    2021

  27. [35]

    Rvae-em: Generative speech dereverberation based on recurrent variational auto-encoder and convolutive transfer function,

    P. Wang and X. Li, “Rvae-em: Generative speech dereverberation based on recurrent variational auto-encoder and convolutive transfer function,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 496–500

    2024

  28. [36]

    Filtered noise shaping for time domain room impulse response estimation from reverberant speech,

    C. J. Steinmetz, V . K. Ithapu, and P. Calamia, “Filtered noise shaping for time domain room impulse response estimation from reverberant speech,” in2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2021, pp. 221–225

    2021

  29. [37]

    Deep impulse responses: Estimating and parameterizing filters with deep networks,

    A. Richard, P. Dodds, and V . K. Ithapu, “Deep impulse responses: Estimating and parameterizing filters with deep networks,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 3209–3213

    2022

  30. [38]

    Unsupervised blind joint dereverberation and room acoustics estimation with diffusion models,

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

    2024 arXiv

  31. [39]

    System identification in the short-time fourier transform domain with crossband filtering,

    Y . Avargel and I. Cohen, “System identification in the short-time fourier transform domain with crossband filtering,”IEEE transactions on Audio, Speech, and Language processing, vol. 15, no. 4, pp. 1305–1319, 2007

    2007

  32. [40]

    Relative transfer function identification using convolutive transfer function approximation,

    R. Talmon, I. Cohen, and S. Gannot, “Relative transfer function identification using convolutive transfer function approximation,”IEEE Transactions on audio, speech, and language processing, vol. 17, no. 4, pp. 546–555, 2009

    2009

  33. [41]

    M. J. Beal,Variational algorithms for approximate Bayesian inference. University of London, University College London (United Kingdom), 2003

    2003

  34. [42]

    The variational ap- proximation for bayesian inference,

    D. G. Tzikas, A. C. Likas, and N. P. Galatsanos, “The variational ap- proximation for bayesian inference,”IEEE Signal Processing Magazine, vol. 25, no. 6, pp. 131–146, 2008

    2008

  35. [43]

    Personal sound zones in the short-time fourier transform domain with relaxed reverberation,

    J. Tang, W. Zhu, and X. Li, “Personal sound zones in the short-time fourier transform domain with relaxed reverberation,”The Journal of the Acoustical Society of America, vol. 157, no. 2, pp. 778–796, 2025

    2025

  36. [44]

    On information and sufficiency,

    S. Kullback and R. A. Leibler, “On information and sufficiency,”The annals of mathematical statistics, vol. 22, no. 1, pp. 79–86, 1951

    1951

  37. [45]

    Nonnegative matrix factor- ization with the itakura-saito divergence: With application to music analysis,

    C. F ´evotte, N. Bertin, and J.-L. Durrieu, “Nonnegative matrix factor- ization with the itakura-saito divergence: With application to music analysis,”Neural computation, vol. 21, no. 3, pp. 793–830, 2009

    2009

  38. [46]

    Machine learning in acoustics: Theory and ap- plications,

    M. J. Bianco, P. Gerstoft, J. Traer, E. Ozanich, M. A. Roch, S. Gannot, and C.-A. Deledalle, “Machine learning in acoustics: Theory and ap- plications,”The Journal of the Acoustical Society of America, vol. 146, no. 5, pp. 3590–3628, 2019

    2019

  39. [47]

    Noise power spectral density estimation based on optimal smoothing and minimum statistics,

    R. Martin, “Noise power spectral density estimation based on optimal smoothing and minimum statistics,”IEEE Transactions on speech and audio processing, vol. 9, no. 5, pp. 504–512, 2001

    2001

  40. [48]

    Usdnet: Unsupervised speech dereverberation via neural forward filtering,

    Z.-Q. Wang, “Usdnet: Unsupervised speech dereverberation via neural forward filtering,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 3882–3895, 2024

    2024

  41. [49]

    Comparison of different impulse response measurement techniques,

    G.-B. Stan, J.-J. Embrechts, and D. Archambeau, “Comparison of different impulse response measurement techniques,”Journal of the Audio engineering society, vol. 50, no. 4, pp. 249–262, 2002

    2002

  42. [50]

    Simultaneous measurement of impulse response and dis- tortion with a swept-sine technique,

    A. Farina, “Simultaneous measurement of impulse response and dis- tortion with a swept-sine technique,” inAudio engineering society convention 108. Audio Engineering Society, 2000

    2000

  43. [51]

    The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,

    C. K. Reddy, V . Gopal, R. Cutler, E. Beyrami, R. Cheng, H. Dubey, S. Matusevych, R. Aichner, A. Aazami, S. Braun, P. Rana, S. Srinivasan, and J. Gehrke, “The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” pp....

    2020

  44. [52]

    Investi- gating rnn-based speech enhancement methods for noise-robust text-to- speech

    C. Valentini-Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “Investi- gating rnn-based speech enhancement methods for noise-robust text-to- speech.” inSSW, 2016, pp. 146–152

    2016

  45. [53]

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

    C. K. Reddy, V . Gopal, and R. Cutler, “Dnsmos p. 835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 886–890

    2022

  46. [54]

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

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

    2024

  47. [55]

    gpurir: A python library for room impulse response simulation with gpu acceleration,

    D. Diaz-Guerra, A. Miguel, and J. R. Beltran, “gpurir: A python library for room impulse response simulation with gpu acceleration,” Multimedia Tools and Applications, vol. 80, no. 4, pp. 5653–5671, 2021

    2021

  48. [56]

    Assessment for automatic speech recog- nition: II. noisex-92: A database and an experiment to study the effect of additive noise on speech recognition systems,

    A. Varga and H. J. Steeneken, “Assessment for automatic speech recog- nition: II. noisex-92: A database and an experiment to study the effect of additive noise on speech recognition systems,”Speech communication, vol. 12, no. 3, pp. 247–251, 1993

    1993

  49. [57]

    The reverb challenge: A common evaluation framework for dereverberation and recognition of reverberant speech,

    K. Kinoshita, M. Delcroix, T. Yoshioka, T. Nakatani, E. Habets, R. Haeb- Umbach, V . Leutnant, A. Sehr, W. Kellermann, R. Maaset al., “The reverb challenge: A common evaluation framework for dereverberation and recognition of reverberant speech,” in2013 IEEE Workshop on Applic...

    2013

  50. [58]

    The design for the wall street journal-based csr corpus,

    D. B. Paul and J. Baker, “The design for the wall street journal-based csr corpus,” inSpeech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992

    1992

  51. [59]

    Estimation of room acoustic parameters: The ace challenge,

    J. Eaton, N. D. Gaubitch, A. H. Moore, and P. A. Naylor, “Estimation of room acoustic parameters: The ace challenge,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 24, no. 10, pp. 1681– 1693, 2016

    2016

  52. [60]

    Multichannel long-term streaming neural speech enhancement for static and moving speakers,

    C. Quan and X. Li, “Multichannel long-term streaming neural speech enhancement for static and moving speakers,”IEEE Signal Processing Letters, vol. 31, pp. 2295–2299, 2024

    2024

  53. [61]

    Cleanmel: Mel-spectrogram enhancement for improving both speech quality and asr,

    N. Shao, R. Zhou, P. Wang, X. Li, Y . Fang, Y . Yang, and X. Li, “Cleanmel: Mel-spectrogram enhancement for improving both speech quality and asr,”IEEE Transactions on Audio, Speech and Language Processing, pp. 1–13, 2025

    2025

  54. [62]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inICLR, 2018

    2018

  55. [63]

    D. M. Howard and D. T. Murphy,Voice science, acoustics, and record- ing. Plural Publishing, 2007

    2007

  56. [64]

    New method of measuring reverberation time,

    M. R. Schroeder, “New method of measuring reverberation time,”The Journal of the Acoustical Society of America, vol. 37, no. 6 Supplement, pp. 1187–1188, 1965

    1965

  57. [65]

    Skipconvnet: Skip convolutional neural network for speech derever- 14 beration using optimally smoothed spectral mapping,

    V . Kothapally, W. Xia, S. Ghorbani, J. H. Hansen, W. Xue, and J. Huang, “Skipconvnet: Skip convolutional neural network for speech derever- 14 beration using optimally smoothed spectral mapping,”arXiv preprint arXiv:2007.09131, 2020

    2007 arXiv

  58. [66]

    Signal estimation from modified short-time fourier transform,

    D. Griffin and J. Lim, “Signal estimation from modified short-time fourier transform,”IEEE Transactions on acoustics, speech, and signal processing, vol. 32, no. 2, pp. 236–243, 1984

    1984

  59. [67]

    NARA- WPE: A python package for weighted prediction error dereverberation in Numpy and Tensorflow for online and offline processing,

    L. Drude, J. Heymann, C. Boeddeker, and R. Haeb-Umbach, “NARA- WPE: A python package for weighted prediction error dereverberation in Numpy and Tensorflow for online and offline processing,” in13. ITG Fachtagung Sprachkommunikation (ITG 2018), Oct 2018

    2018

  60. [68]

    Blind estimation of the coherent-to-diffuse energy ratio from noisy speech signals,

    M. Jeub, C. Nelke, C. Beaugeant, and P. Vary, “Blind estimation of the coherent-to-diffuse energy ratio from noisy speech signals,” in2011 19th European Signal Processing Conference. IEEE, 2011, pp. 1347–1351

    2011

  61. [69]

    Per- ceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Per- ceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,” in2001 IEEE international conference on acoustics, speech, and signal processing. Procee...

    2001

  62. [70]

    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 Transac- tions on Audio, Speech, and Language Processing, vol. 24, no. 11, pp. 2009–2022, 2016

    2009

  63. [71]

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

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

    2021

  64. [72]

    A speech enhancement approach using piecewise linear approximation of an explicit model of environmental distortions

    J. Du and Q. Huo, “A speech enhancement approach using piecewise linear approximation of an explicit model of environmental distortions.” inInterspeech, 2008, pp. 569–572

    2008

  65. [73]

    ITU-T Recommendation P.808, Subjective evaluation of speech quality with a crowdsourcing approach,

    International Telecommunication Union, “ITU-T Recommendation P.808, Subjective evaluation of speech quality with a crowdsourcing approach,” Geneva, 2018

    2018

  66. [74]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  67. [75]

    Investigation into joint optimization of single channel speech enhancement and acoustic modeling for robust asr,

    T. Menne, R. Schl ¨uter, and H. Ney, “Investigation into joint optimization of single channel speech enhancement and acoustic modeling for robust asr,” inICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 6660–6664

    2019

  68. [76]

    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 Acous- tics, vol. 19, no. 1. Acoustical Society of America, 2013, p. 035081

    2013

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.