Pith. sign in

REVIEW 4 major objections 8 minor 34 references

Speech Enhancement with Dual-path Multi-Channel Linear Prediction Filter and Multi-norm Beamforming

T0 review · 4 major / 8 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that combining temporal and frequential MCLP filters with an l1-regularized beamformer outperforms standard cascades in reverberant speech enhancement.

desk verdict A plausible dual-path MCLP + l1-beamforming combination with a nice order-selection idea, but the complex soft-thresholding is ambiguous and the per-T60 order choices look tuned to the test PESQ curves. read the letter →

arxiv 2507.18350 v1 pith:MV5IZWNJ submitted 2025-07-24 eess.AS cs.SD

Chengyuan Qin , Wenmeng Xiong , Jing Zhou , Maoshen Jia , Changchun Bao This is my paper · ORCID
classification eess.AScs.SD
keywords speechenhancementmultichannellinearpredictiondereverberationbeamformingl1sparsityorderselectionmicrophonearraydual-pathfiltering
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

The paper sets out to show that microphone-array speech enhancement in noisy, reverberant rooms improves when late reflections are predicted along both the time axis and the frequency axis of the STFT domain, rather than time alone, and when the subsequent beamformer minimizes both output power and the $\ell^1$ norm of its output. It combines a dual-path multichannel linear prediction (MCLP) filter with a multi-norm beamformer and evaluates the pair on simulated rooms. The reported result is higher PESQ and SI-SNR than the GWPE, GWPE+MVDR, and WPD baselines, with the largest advantage at high reverberation (T60 of 0.4 to 1.0 s). The paper also offers a correlation-threshold rule for choosing prediction orders that is meant to transfer to other MCLP methods. If correct, the method is a training-free signal-processing route to better far-field speech quality, and a simple knob for tuning dereverberation filters.

What carries the argument

The load-bearing mechanism is a pair of filter matrices: G_t, a temporal MCLP filter estimated per frequency bin, and G_f, a frequential filter estimated per time frame, which together predict late reverberation from stacked observations in both directions. The filters are found by minimizing a summed $\ell^2$ plus $\ell^1$ cost over dereverberated STFT coefficients via Proximal Alternating Linearized Minimization (PALM), with soft thresholding supplying the $\ell^1$ proximal step. A second stage applies a multi-norm beamformer—output power plus $\ell^1$ penalty under a distortionless constraint—solved by ADMM. Prediction orders K_t and K_f are set by thresholding Pearson correlation coefficients between reference-microphone samples at increasing time or frequency lags, replacing grid search. The $\ell^1$ terms are the reason the method is called 'multi-norm': both stages mix power ($\ell^2$) minimization with sparsity ($\ell^1$) regularization.

What would settle it

Re-solve Eq. (5) with a proximal update that shrinks the magnitude of each complex STFT coefficient while keeping its phase, then rerun the reported TIMIT experiments; if PESQ and SI-SNR no longer beat the baselines, the claimed gains depend on the incorrectly specified thresholding step rather than on the dual-path or multi-norm idea.

Watch

Extended reading notes

Core claim

The central claim is that jointly estimating two MCLP filter matrices—one that operates across time frames at each frequency and one that operates across frequency bins at each time frame—removes late reverberation more completely than temporal-only prediction, and that adding an $\ell^1$ sparsity penalty to a distortionless beamformer's $\ell^2$ power cost improves denoising. On 8-microphone simulated arrays with TIMIT speech and image-method room responses, the paper reports that this dual-path, multi-norm system outperforms GWPE, GWPE+MVDR, and WPD on PESQ and SI-SNR for T60 from 0.4 to 1.0 s and across all tested SNRs, with the gains concentrated in heavy reverberation. The authors attribute the improvement to more comprehensive modeling of late reverberation by the frequential filter path and to the sparsity prior on the enhanced output.

Load-bearing premise

The load-bearing premise is that the soft-thresholding step used to enforce sparsity is valid for the complex STFT coefficients it is applied to, but the paper defines the operator only for real scalars; if an implementation applies that formula directly to complex values, the algorithm no longer solves the stated $\ell^2$+$\ell^1$ problem.

Editorial extensions

If this is right

  • At T60 values from 0.4 to 1.0 s, the proposed method reports higher PESQ and SI-SNR than GWPE, GWPE+MVDR, and WPD, and the margin grows with reverberation.
  • The l1 norm on beamformer output adds denoising power beyond power minimization, yielding gains over WPD across all tested SNR levels at moderate T60.
  • The Pearson-correlation threshold method selects temporal and frequential prediction orders without grid search and carries over to other MCLP-based systems.
  • At very low T60 (0.2 s) the temporal-only WPD baseline is still slightly better, implying a T60-aware switch between dual-path and temporal-only filtering would be useful.

Reading between the lines

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

  • The complex soft-thresholding inconsistency in Eq. (9) and Eq. (14) means the reported gains should be rechecked with a magnitude-based complex proximal operator; the l1 objective may not be what is actually minimized.
  • Because the experiments use Gaussian white noise and simulated RIRs, the method's advantage in real rooms with babble or diffuse noise is unverified; a test with measured RIRs and nonstationary noise would be the natural next check.
  • The order-selection recipe could be plugged into standard WPE or MCLP pipelines as a cheap heuristic for choosing filter lengths, which would make the paper's contribution useful even if the dual-path gains do not replicate.
  • The sparsity penalty on beamformer output may trade off intelligibility against perceived quality; reporting STOI or word-error-rate on a downstream recognizer would clarify whether the PESQ gains are practically useful.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper proposes a speech enhancement algorithm that combines dual-path multi-channel linear prediction (MCLP) filters, operating in both time and frequency dimensions, with a minimum-power distortionless beamformer that also penalizes the l1 norm of the output. An auxiliary contribution is a method for selecting MCLP prediction orders based on Pearson correlation thresholds. The method is evaluated in simulated reverberant environments using TIMIT speech, an 8-microphone uniform linear array, and compared against GWPE, GWPE+MVDR, and WPD using PESQ and SI-SNR. The reported results show advantages over the baselines especially at long T60 values and are claimed to support the proposed order-selection rule.

Significance. If the technical issues are resolved, the dual-path extension of MCLP to frequency-domain prediction is a plausible and interesting direction, and a validated data-driven order-selection rule would be practically useful for MCLP-based dereverberation methods. However, the current manuscript contains two load-bearing problems: the complex soft-thresholding update is incorrectly specified, and the order-selection rule is not demonstrably the source of the experimental settings. These must be addressed before the reported performance gains can be attributed to the proposed method.

major comments (4)
  1. [Sec. 3.1.2, Eq. (9)] The soft-thresholding operator S_{λ_z/μ_z}(v) is defined for real scalars via two one-sided inequalities, but it is applied to complex STFT vectors z(n,ω). The correct proximal map for the complex ℓ1 norm is element-wise magnitude shrinkage, i.e., S_τ(v) = max(|v|−τ,0) · v/|v|, which is neither stated nor implied by the displayed definition. As written, the z-update in Eq. (9) does not implement the proximal step for the ℓ1 term in Eq. (5), and the same issue affects the complex scalar update in Eq. (14). Please correct the operator definition and confirm that the implementation uses the complex shrinkage form.
  2. [Secs. 4 and 5] The prediction-order selection method is not connected to the experiments. Equation (18) depends on thresholds δ1 and δ2, but no numerical values are given for these thresholds. In Sec. 5, the paper lists K_t = {10,14,18,22,24} and K_f = {2,4,6,8,10} for T60 = {0.2,...,1.0} s but does not state that these values were obtained from Eq. (18). Since Fig. 1 displays PESQ curves for K_t under each T60 and marks the selected orders, it is unclear whether the orders used in Fig. 2 were chosen from the PESQ curves themselves, which would bias the comparison in favor of the proposed method and would not demonstrate the advertised order-selection contribution. Please specify δ1 and δ2, present the K_t and K_f values predicted by Eq. (18) for each T60, and compare them with the values used in the simulations.
  3. [Sec. 5, Fig. 2] The paper reports mean PESQ and SI-SNR over 100 Monte Carlo runs but gives no measure of variability. Several reported differences are small; for example, at T60 = 0.2 s the proposed method is below WPD in PESQ, and at higher SNRs some gaps are within a few hundredths of a point. Without standard deviations, confidence intervals, or significance tests, the reader cannot assess whether the observed ordering of methods is statistically meaningful. Please add error bars or significance tests, and clarify whether the same noise and reverberation realizations are used for all methods in each Monte Carlo run.
  4. [Sec. 4, Eq. (17)] Equation (17) defines the Pearson correlation coefficient across Monte Carlo realizations i between values at time indices 0 and t. This is an ensemble correlation, not the temporal autocorrelation that is conventionally used for prediction-order selection. If the intended measure is the sample autocorrelation of a single recording, the equation and the surrounding description need to be corrected. If the ensemble correlation is truly intended, its relationship to the optimal MCLP prediction order should be justified. Please clarify this point, as it is load-bearing for the order-selection method.
minor comments (8)
  1. [Abstract] There are stray spacing issues in the abstract, e.g., 'us ing' and 'thel 1' should be 'using' and 'the l1'.
  2. [Sec. 2.1] The phrase 'Korder convolution' should be 'K-th order convolution'.
  3. [Sec. 4] The sentence 'The prediction order selection method is present' should be 'is presented'.
  4. [Fig. 1] The caption 'Pearson correlation coefficients and PESQ with different T60 values' is vague; please label the axes and panel subcaptions to make clear which quantity is plotted and which T60 applies to each panel.
  5. [Sec. 5] The text 'the additive noises are Gaussian white' should be 'the additive noise is white Gaussian noise'.
  6. [Sec. 5] The experimental setup does not specify the number of sources Q, the locations of the target source and noise sources, or whether the noise is diffuse or a point source. Please state these details, as they affect the interpretation of the beamforming results.
  7. [Secs. 3 and 5] The hyperparameters λ_z, λ_w, ρ_G, ρ_w, μ_z, μ_w, γ, γ_w, and γ_1 are not given values in the text. A table with the selected values and the tuning procedure would improve reproducibility.
  8. [Conclusion] The claim that the proposed order-selection method 'can also be applied to other MCLP-based methods' is not supported by any experiment; please either add such an experiment or temper the claim.

Circularity Check

1 steps flagged · score 2.0 of 10

The order-selection thresholds are calibrated on the proposed method's own PESQ curves, making the claimed robust selection and high-T60 advantage partly fitted rather than independently predicted.

  1. fitted input called prediction [Section 4, Eq. (18) and Fig. 1; Section 5, Fig. 2]
    "In our experiments, we have found that, for different T60 configurations, a single threshold δ can be selected within the range δ1≤δ≤δ2, such that the corresponding time lag is also an approximate optimal prediction order of the temporal filters in our proposed MCLP based method. ... The prediction orders corresponding to the rose-red circles of PESQ values in Fig. 1(b)-(d) are equal to the time lags corresponding to δ1 and δ2 in Fig. 1(a). Hence, in our following work, the optimal prediction order in the time filters is selected as: Kt = 1/2(Kδ1 + Kδ2),"

    The thresholds δ1 and δ2 are not set by an independent physical or statistical criterion: the paper says the lags matching them are equal to the PESQ-optimal K_t values of the proposed method. Thus Eq. (18) selects K_t by reading off the proposed method's own test PESQ curve, so the claimed robust, T60-adaptive order selection is a fit to the performance surface it is supposed to predict. Section 5 then reports K_t={10,14,18,22,24} and K_f={2,4,6,8,10} for the same T60 values without showing they were produced by Eq. (18) from Pearson correlations alone, so the high-reverberation gains in Fig. 2 may reflect favorable order choices rather than an independent prediction. The core Sec. 3 optimization is not affected.

full rationale

The main signal-processing derivation is self-contained: Sec. 3 minimizes explicitly stated objectives, Eqs. (5) and (12), and the comparisons in Fig. 2 are against external baselines (GWPE, GWPE+MVDR, WPD), so the central enhancement claim is not circular. The only circular element is the order-selection method: Fig. 1 calibrates δ1 and δ2 using the proposed method's own PESQ-versus-K_t curves on the same T60 conditions, and Eq. (18) converts those lags into the selected orders. Because Sec. 5 lists the K_t and K_f values without demonstrating that Eq. (18) produced them on independent data, the robustness and high-reverberation advantage are partly fitted rather than independently predicted. The real-valued soft-threshold operator applied to complex vectors in Eqs. (9) and (14) is a correctness and reproducibility concern, but not a circularity concern.

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

The method rests on the standard MCLP signal model and sparsity assumptions, plus a heuristic order-selection threshold that is tuned to the evaluated room conditions. No new physical entities are introduced. The many undisclosed optimization hyperparameters increase the effort needed to reproduce the results.

free parameters (5)
  • λ_z = not given
    Sparsity penalty weight in the dual-path MCLP objective (Eq. 5); chosen by hand, no value or sensitivity analysis provided.
  • λ_w = not given
    Sparsity penalty weight in the beamforming objective (Eq. 12); chosen by hand, no value or sensitivity analysis provided.
  • δ1, δ2 (correlation thresholds) = not given
    Thresholds used in the prediction order selection method (Sec. 4); selected empirically so that the resulting time lag matches the PESQ-optimal order in Fig. 1, which tunes them to the evaluated conditions.
  • K_t, K_f (prediction orders) = K_t = {10,14,18,22,24}, K_f = {2,4,6,8,10} for T60 = {0.2,0.4,0.6,0.8,1.0}s
    Prediction orders for time and frequency filters; set per T60 using the proposed correlation method and used in the main evaluation (Sec. 5), meaning the experimental hyperparameters are tuned to the test conditions.
  • ADMM/PALM hyperparameters (ρ_G, ρ_w, μ_z, μ_w, γ, γ_w, γ_1) = not given
    Penalty and step-size parameters in the iterative solvers (Eqs. 6-10, 13-15, 20-21); no values or convergence checks reported.
assumptions (5)
  • domain assumption The TF-domain convolution model in Eq. (1) accurately represents microphone array signals in noisy and reverberant rooms.
    All derivations start from this model (Talmon et al. [20]); if real RIRs deviate significantly, the MCLP decomposition fails.
  • domain assumption Speech STFT coefficients are sparse, justifying the l1 norm penalties in Eq. (5) and (12).
    Used to motivate the multi-norm objectives; sparsity of speech TF coefficients is a standard but not universally valid assumption, especially for multi-talker or noisy conditions.
  • ad hoc to paper A single correlation threshold δ (with δ1 ≤ δ ≤ δ2) maps to near-optimal prediction orders across different T60 values.
    This is an empirical claim in Sec. 4 supported only by Fig. 1 for three T60 values and the paper's own method; no theoretical justification or cross-validation is given.
  • domain assumption The steering vector a(θ_s) of the target source is known exactly for the distortionless constraint in Eq. (12).
    The beamforming constraint assumes perfect DOA information; estimation errors are not considered.
  • domain assumption Simulated RIRs (image method) and Gaussian white noise are representative of real-world acoustic environments.
    The evaluation is entirely simulated (Sec. 5); generalization to real rooms and non-white noises is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Speech Enhancement with Dual-path Multi-Channel Linear Prediction Filter and Multi-norm Beamforming." pith.science (2026). https://pith.science/paper/MV5IZWNJ

@misc{pith2026250718350,
  author       = {Pith},
  title        = {Pith review of: Speech Enhancement with Dual-path Multi-Channel Linear Prediction Filter and Multi-norm Beamforming},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MV5IZWNJ}},
  note         = {Machine review of arXiv:2507.18350}
}
read the original abstract

In this paper, we propose a speech enhancement method us ing dual-path Multi-Channel Linear Prediction (MCLP) filters and multi-norm beamforming. Specifically, the MCLP part in the proposed method is designed with dual-path filters in both time and frequency dimensions. For the beamforming part, we minimize the power of the microphone array output as well as the l1 norm of the denoised signals while preserving source sig nals from the target directions. An efficient method to select the prediction orders in the dual-path filters is also proposed, which is robust for signals with different reverberation time (T60) val ues and can be applied to other MCLP-based methods. Eval uations demonstrate that our proposed method outperforms the baseline methods for speech enhancement, particularly in high reverberation scenarios.

Figures

Figures reproduced from arXiv: 2507.18350 by the authors.

Figure 1
Figure 1. Pearson correlation coefficients and PESQ with dif￾ferent T60 values. 4. Proposed Prediction Order Selection Method In this section, the prediction order selection method is present. Firstly, let us define y i 1t (t), t = 0, ..., T as the i th individual sample of the temporal signals at the reference microphone at time index t in the Monte Carlo experiments. The Pearson cor￾relation coefficients between y i 1t (0) … view at source ↗
Figure 2
Figure 2. (c) and 2(d) show PESQ and SI-SNR as a function of SNR, with T60 set to 0.3s . Kt = 12 and Kf = 3. It can be ob￾served that, as a joint optimization method, WPD outperforms the cascaded GWPE and MVDR method (”GWPE+MVDR”). However, due to the l1 norm constraint on the denoised signals, our proposed method exhibits a more powerful denoising capa￾bility and outperforms all the baseline methods across all the SNRs. 6. C… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

  1. [1]

    Introduction Speech enhancement technique is of great importance for nu- merous applications such as automatic speech recognition, human-machine interaction, and smart home devices[1, 2, 3]. Conventional speech enhancement methods, including spectral subtraction [4], wiener filtering [5], and subspace-based meth- ods [6], have been widely investigated and...

    work page

  2. [2]

    Speech Enhancement with Dual-path Multi-Channel Linear Prediction Filter and Multi-norm Beamforming

    Preliminaries 2.1. Microphone array signal model Considering thatQfar-field wideband acoustic sources impinge onMmicrophones in a noisy and reverberant room. The sig- nals received at the microphone array in the TF domain are ap- proximately formulated as aKorder convolution between the STFT of the room impulse response (RIR)h q(n,ω)and the STFT signals q...

    work page Pith review arXiv 2025

  3. [3]

    Proposed Algorithm The proposed algorithm consists of two parts: dual-path MCLP filters for dereverberation and multi-norm beamforming for de- noising. 3.1. Dual-path MCLP filter branch 3.1.1. Dereverberated signals with dual-path filters Let us defineG f(n) = [g f1(n),...,g fm(n),...,g fM (n)]∈ C((2Kf +1)M)×M as the frequential filter matrix at time fram...

    work page

  4. [4]

    Firstly, let us definey i 1t(t),t= 0,...,Tas thei th individual sample of the temporal signals at the reference microphone at time indextin the Monte Carlo experiments

    Proposed Prediction Order Selection Method In this section, the prediction order selection method is present. Firstly, let us definey i 1t(t),t= 0,...,Tas thei th individual sample of the temporal signals at the reference microphone at time indextin the Monte Carlo experiments. The Pearson cor- relation coefficients betweenyi 1t(0)andy i 1t(t)can be given...

    work page

  5. [5]

    GWPE”) , the cascade of GWPE and MVDR method [12] (legend: “GWPE+MVDR

    Simulation Experiments In this section, numerical simulations are presented to illus- trate the validity of the proposed method. Source signals are speeches from the TIMIT database [25] sampled at16kHz. A uniform linear microphone array (ULA) composed of8micro- phones with the inter-element space equal to0.03m is utilised. The RIRs are generated using the...

    work page

  6. [6]

    The pro- posed method demonstrates superior performance in both dere- verberation and denoising compared to the baseline methods, particularly in high reverberation scenarios

    Conclusion In this paper, we propose a speech enhancement method using dual-path MCLP filters and multi-norm beamforming. The pro- posed method demonstrates superior performance in both dere- verberation and denoising compared to the baseline methods, particularly in high reverberation scenarios. In addition, we have proposed an efficient method for selec...

    work page

  7. [7]

    Appendix To solvew (l+1), a new augmented Lagrangian can be derived as: L(w(l+1),η 1) = NX n=1 (∥w(l+1)Hˆx(n,ω)∥ 2 2 +Re{η (l)H w (w(l+1)H ˆx(n,ω)−z (l) w (n,ω))}+ 1 2ρw ∥w(l+1)Hˆx(n,ω)−z (l) w (n,ω)∥ 2 2)+ Re{ηH 1 (w(l+1)H a(θs)−1)}+ 1 2ρ1 ∥w(l+1)H a(θs)−1∥ 2 2, (19) The problem (19) can be solved via several iterative steps. In the(j+ 1) th iteration,wc...

    work page

  8. [8]

    Acknowledgements This work was supported by the National Natural Science Foun- dation of China (Grant No.62101013)

    work page

Show all 34 references
  1. [9]

    Prediction of speech intelligibility in spatial noise and reverberation for normal-hearing and hearing- impaired listeners,

    R. Beutelmann and T. Brand, “Prediction of speech intelligibility in spatial noise and reverberation for normal-hearing and hearing- impaired listeners,”J. Acoust. Soc. Amer, vol. 120, no. 1, pp. 331– 342, Jul. 2006

    2006

  2. [10]

    Environmental con- ditions and acoustic transduction in hands-free speech recogni- tion,

    M. Omologo, P. Svaizer, and M. Matassoni, “Environmental con- ditions and acoustic transduction in hands-free speech recogni- tion,”Speech Communication, vol. 25, no. 1-3, pp. 75–95, Aug. 1998

    1998

  3. [11]

    On the application of reverberation suppression to robust speech recog- nition,

    R. Maas, E. A. Habets, A. Sehr, and W. Kellermann, “On the application of reverberation suppression to robust speech recog- nition,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Kyoto, Japan, Mar. 2012, p. 297–300

    2012

  4. [12]

    Ephraim and I

    Y . Ephraim and I. Cohen,Recent advancements in speech en- hancement. The Electronic Handbook, 2006

    2006

  5. [13]

    Interaural coherence preservation in multi-channel wiener filtering based noise reduc- tion for binaural hearing aids,

    D. Marquardt, V . Hohmann, and S. Doclo, “Interaural coherence preservation in multi-channel wiener filtering based noise reduc- tion for binaural hearing aids,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, no. 12, p. 2162–2176, Dec. 2015

    2015

  6. [14]

    Improved subspace-based single- channel speech enhancement using generalized super-gaussian priors,

    J. Jensen and R. Heusdens, “Improved subspace-based single- channel speech enhancement using generalized super-gaussian priors,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 15, no. 3, pp. 862–872, Mar. 2007

    2007

  7. [15]

    Precise dereverbera- tion using multichannel linear prediction,

    M. Delcroix, T. Hikichi, and M. Miyoshi, “Precise dereverbera- tion using multichannel linear prediction,”IEEE/ACM Transac- tions on Audio, Speech, and Language Processing, vol. 15, no. 2, p. 430–440, Feb. 2007

    2007

  8. [16]

    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/ACM Transactions on Audio, Speech, and Language Processing, vol. 18, no. 7, p. 1717–1731, Sep. 2010

    2010

  9. [17]

    Generalization of multi-channel linear prediction methods for blind mimo impulse respons short- ening,

    T. Yoshioka and T. Nakatani, “Generalization of multi-channel linear prediction methods for blind mimo impulse respons short- ening,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 20, no. 10, p. 2707–2720, Dec. 2012

    2012

  10. [18]

    Benesty and J

    J. Benesty and J. Chen,Microphone Array Signal Processing. Germany: Springer-Verlag, 2008

    2008

  11. [19]

    Strategies for distant speech recognition in rever- berant environments,

    M. Delcroix, “Strategies for distant speech recognition in rever- berant environments,” no. 1, p. 60, Aug. 2015

    2015

  12. [20]

    Dereverberation with differential micro- phone arrays and the weighted-prediction-error method,

    W. Yang and G. Huang, “Dereverberation with differential micro- phone arrays and the weighted-prediction-error method,” inInter- national Workshop on Acoustic Signal Enhancement(IWAENC), 2018

    2018

  13. [21]

    Multichannel online speech dereverberation under noisy environments,

    M. Togami, “Multichannel online speech dereverberation under noisy environments,” pp. 1078–1082, Aug. 2015

    2015

  14. [22]

    A unified convolutional beam- former for simultaneous denoising and dereverberation,

    T. Nakatani and K. Kinoshita, “A unified convolutional beam- former for simultaneous denoising and dereverberation,”IEEE Signal Process, p. 903–907, Jun. 2019

    2019

  15. [23]

    Dual-path rnn: Efficient long sequence modeling for time-domain single-channel speech sepa- ration,

    Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path rnn: Efficient long sequence modeling for time-domain single-channel speech sepa- ration,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Proc, Oct. 2020, p. 46–50

    2020

  16. [24]

    Dpt-fsnet:dual-path trans- former based full-band and sub-band fusion network for speech enhancement,

    F. Dang, H. Chen, and P. Zhang, “Dpt-fsnet:dual-path trans- former based full-band and sub-band fusion network for speech enhancement,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Singapore, Mar. 2022, p. 6857–6861

    2022

  17. [25]

    Tf-gridnet: Making time- frequency domain models great again for monaural speaker sepa- ration,

    Z. Wang, S. Cornell, and S. Choi, “Tf-gridnet: Making time- frequency domain models great again for monaural speaker sepa- ration,”arXiv preprint arXiv, p. 2209.03952, 2022

    2022 arXiv

  18. [26]

    Spatialnet: Extensively learning spatial in- formation for multichannel joint speech separation, denoising and dereverberation,

    C. Quan and X. Li, “Spatialnet: Extensively learning spatial in- formation for multichannel joint speech separation, denoising and dereverberation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 1310–1323, 2024

    2024

  19. [27]

    Dasformer: Deep alternating spectrogram transformer for multi/single-channel speech separation,

    S. Wang, X. Kong, X. Peng, H. Movassagh, V . Prakash, and Y . Lu, “Dasformer: Deep alternating spectrogram transformer for multi/single-channel speech separation,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), Island, Jun. 2023, pp. 1–5

    2023

  20. [28]

    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/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 17, p. 546–555, May. 2009

    2009

  21. [29]

    Proximal alternating lin- earized minimization for nonconvex and nonsmooth problems,

    J. Bolte, S. Sabach, and M. Teboulle, “Proximal alternating lin- earized minimization for nonconvex and nonsmooth problems,”J. Acoust. Soc. Amer, vol. 146, no. 1/2, p. 459–494, 2014

    2014

  22. [30]

    D. P. Bertsekas,Constrained Optimization and Lagrange Multi- plier Methods. New York: Academic, 1982

    1982

  23. [31]

    An iterative thresh- olding algorithm for linear inverse problems with a sparsity con- straint,

    I. Daubechies, M. Defrise, and C. D. Mol, “An iterative thresh- olding algorithm for linear inverse problems with a sparsity con- straint,”Comm. Pure Appl. Math, vol. 57, no. 11, pp. 1413–1457, 2004

    2004

  24. [32]

    Dis- tributed optimization and statistical learning via the alternating di- rection method of multipliers,

    S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Dis- tributed optimization and statistical learning via the alternating di- rection method of multipliers,”Found. Trends Mach. Learn, no. 1, pp. 1–122, 2011

    2011

  25. [33]

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, and D. S. Pallett,DARPA TIMIT acoustic-phonetic continous speech corpus CD-ROM, 1993

    1993

  26. [34]

    Room impulse response generator,

    E. A. P. Habets, “Room impulse response generator,” in The Netherlands, Tech. Rep. 2.4, 2006. [Online]. Avail- able: https://www.audiolabs-erlangen.de/fau/professor/ habets/software/rir-generator, 2006

    2006

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.