Pith. sign in

REVIEW 2 major objections 5 minor 41 references

Pseudo Labels-based Neural Speech Enhancement for the AVSR Task in the MISP-Meeting Challenge

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A neural speech enhancer trained on close-talk pseudo labels cuts meeting speech recognition errors to 5.44% character error rate, a 64.8% relative gain over the challenge baseline.

desk verdict Solid challenge system, but the ablation doesn't isolate the pseudo-label contribution; the S3-vs-S1 claim needs a control condition. read the letter →

arxiv 2505.24446 v2 pith:WJOTNJVW submitted 2025-05-30 cs.SD eess.AS

classification cs.SDeess.AS
keywords MISP-MeetingChallengespeechenhancementpseudolabelsguidedsourceseparationfar-fieldrecognitiontimealignmentaudio-visualParaformer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a speech enhancement model can be trained directly on real far-field meeting recordings, without simulated data, by using pseudo labels derived from close-talk microphones. The proposed TLS framework aligns the close-talk signal in time with the far-field mixture using GCC-PHAT, then fits a short linear filter to match its level and spectrum, yielding a clean reference for the GSS-enhanced mixture. A neural enhancer, G-SpatialNet, trained on these pairs with a magnitude-aware loss, removes residual noise and reverberation from GSS outputs. On the MISP-Meeting benchmark the full system reaches character error rates of 5.44% on Dev and 9.52% on Eval, a relative improvement of 64.8% and 52.6% over the baseline.

What carries the argument

The argument rests on three pieces: TLS, G-SpatialNet, and the magnitude constraint adjustable (MCA) loss. TLS first uses GCC-PHAT to estimate the time offset between a close-talk segment and the far-field GSS reference, then fits a short multi-frame linear filter per frequency to match the close-talk signal's level and spectral envelope to the reference, and finally discards pairs whose estimated SNR is below -10 dB. G-SpatialNet is SpatialNet reworked to consume stacked magnitude spectrograms of the GSS signal and the raw 8-channel array, with an ideal amplitude mask as the training target. The MCA loss linearly combines MSE with cosine similarity between oracle and predicted magnitude spectrograms, the cosine term acting as a softer constraint that stabilizes training on real data.

What would settle it

Train the same G-SpatialNet architecture on standard simulated noisy-clean pairs and compare it against the pseudo-label-trained model on the same MISP-Meeting test; if simulated training matches or beats pseudo-label training, the claim that real-data pseudo labels are the engine of the gain loses support. In a dataset with known clean references, measure the residual misalignment between TLS pseudo labels and the true clean speech; if the residual offset is large or the spectral mismatch perceptible, the proxy assumption is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that signal-level pseudo labels, produced by aligning close-talk recordings to far-field GSS mixtures, let a neural speech enhancer be trained directly on real meeting audio and thereby improve downstream ASR beyond what GSS alone achieves. With Paraformer as the back-end, swapping GSS output for G-SpatialNet output lowers Dev CER from 6.31% to 6.01%; feeding both GSS and the raw 8-channel array into the enhancer brings it to 5.46% Dev and 9.58% Eval, a relative gain of 13.5% and 12.1% over GSS-only. Fine-tuning Paraformer in two stages, first on pseudo labels and then on G-SpatialNet-enhanced audio, yields the final system at 5.44% Dev and 9.52% Eval, which is 64.8% and 52.6% better than the challenge baseline. The authors also find that audio-only fine-tuning beats audio-visual fine-tuning and that additional external meeting data does not improve the pre-trained model.

Load-bearing premise

The load-bearing premise is that after GCC-PHAT time alignment and multi-frame linear filtering, the filtered close-talk signal is a faithful stand-in for the clean speech hidden inside the GSS mixture, so any timing or spectral error baked into those training targets transfers to the enhanced output.

Editorial extensions

If this is right

  • Training speech enhancement on real far-field data via pseudo labels is a viable path when clean references do not exist, removing the simulated-to-real domain gap that degrades conventional SE models.
  • Feeding both GSS and raw multi-channel array input to the enhancer yields further ASR gains, attributed to the reuse of spatial information.
  • Fine-tuning a large pre-trained ASR on the enhancer's output and on pseudo labels aligns front-end and back-end, producing a small but consistent additional CER reduction.
  • Adding more meeting data (AliMeeting, AISHELL-4, and simulated data) did not improve a 60,000-hour pre-trained Paraformer, while audio-only fine-tuning beat audio-visual fine-tuning on this task.
  • The SNR filter that removes pairs below -10 dB improves training stability, indicating that data selection matters for pseudo-label-based training.

Reading between the lines

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

  • The same TLS recipe could be applied to other meeting corpora that record both close-talk and far-field microphones, adapting a general enhancer to a target room with no simulated data at all.
  • Because the level-alignment filter in Eq. 2 is fitted to the GSS signal itself, the pseudo labels may inherit GSS's spectral colouring; an iterative scheme that feeds enhanced output back as a new reference could decouple the enhancer from initial GSS artifacts.
  • The -10 dB SNR threshold is a coarse heuristic; a soft weighting of the loss by estimated SNR might let the model use more data without letting low-quality pairs destabilize training.
  • The result suggests that for very large pre-trained ASR models, target-domain fine-tuning on the actual enhanced audio matters more than adding external data, a recipe that could be tested on other distant-talk benchmarks.
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 describes the authors' system for the MISP-Meeting Challenge Track 2 (AVSR). The front-end consists of Guided Source Separation (GSS) followed by a neural speech enhancement model called G-SpatialNet, an adaptation of SpatialNet that takes both the GSS signal and the raw 8-channel array as input. To train this model on real far-field data without clean references, the paper proposes TLS (time alignment, level alignment, SNR filtering) to generate pseudo labels from close-talk segments, together with a magnitude-constraint adjustable (MCA) loss. The back-end is Paraformer-large, fine-tuned in two stages on pseudo labels and G-SpatialNet-enhanced audio. The system achieves 5.44% CER on Dev and 9.52% on Eval, a relative improvement of 64.8% and 52.6% over the challenge baseline, placing second. The main claim is that the pseudo-label-based training of G-SpatialNet yields the gains observed when moving from system S1 (GSS only) to S3 (GSS + G-SpatialNet with 8-channel input).

Significance. If the TLS pseudo-label pipeline is shown to be responsible for the gains, the paper would make a useful contribution: it would provide a recipe for training neural speech enhancement directly on real meeting recordings, avoiding the simulated-to-real domain mismatch that is a known limitation in distant speech recognition. The reported challenge results are strong and the Dev/Eval numbers are externally evaluated. The paper also includes a useful negative result on data augmentation for a large pre-trained ASR model and on the visual modality. The supporting evidence for the central attribution, however, is not yet conclusive: the ablations in Table 3 do not isolate the pseudo-label contribution from the introduction of the enhancement architecture and the 8-channel spatial input. The paper would be strengthened by a control condition in which the same G-SpatialNet architecture is trained on simulated data with known clean targets, or on raw close-talk targets without TLS, so that the specific benefit of the pseudo labels can be separated from the benefit of adding a learned enhancement stage.

major comments (2)
  1. [Section 5.3, Table 3] The central claim, stated in Section 5.3 as 'system S3 achieves relative performance improvements of 13.5% and 12.1% over system S1 ..., which fully demonstrates the effectiveness of our proposed pseudo labels-based neural speech enhancement method,' is not supported by the reported ablation design. Moving from S1 to S3 changes three factors at once: (i) a neural enhancement network is inserted, (ii) the network receives the raw 8-channel array in addition to the GSS signal, and (iii) the network is trained on TLS pseudo labels. The comparison S2 vs. S1 shows that adding G-SpatialNet trained on pseudo labels (without the 8-channel input) yields a smaller gain (6.31% to 6.01% on Dev), while comparing S2 vs. S3 shows that adding the 8-channel input yields a larger gain (6.01% to 5.46% on Dev). Without a control condition that trains the same G-SpatialNet architecture on simulated data with known clean targets or on close-talk targets with only trivial alignment, the observed S1-to-S3 improvement cannot be attributed specifically to the pseudo-label pipeline. The sentence beginning 'fully demonstrates' is therefore an overclaim relative to the evidence.
  2. [Section 2.2.2, Eq. (2)] The level-alignment step fits a short multi-frame linear filter h'(f) by minimizing the difference between the filtered close-talk signal and the GSS output y(t), where y(t) is explicitly designated as the 'Reference Far-Field Mixture' and is later used as the input to G-SpatialNet. This makes the pseudo-label target correlated with the network input by construction. Because the minimization uses the same y(t) as the reference, a short filter could partially bias the pseudo label toward the noisy/reverberant characteristics of y(t), making the learning task easier without guaranteeing that the pseudo labels represent the true clean speech component. The paper currently validates the pseudo labels only indirectly via CER and DNSMOS in Table 1, not by any direct alignment-error or speech-distortion measurement. As a consequence, Eq. (2) leaves open the possibility that the improvement attributed to pseudo labels is actually due to the network learning a partially identity-like mapping. A control with simulated mixtures where the true clean reference is known, or a direct measurement of the alignment residual, would address this concern.
minor comments (5)
  1. [Abstract and throughout] The text contains a repeated spacing error: 'A VSR' should be 'AVSR' in the abstract and in Section 1.
  2. [Section 5 heading] 'Result and analysize' should be 'Results and Analysis'.
  3. [Section 2.2.2] The phrase 'an time offset' should be 'a time offset', and 'an level-misalignment' should be 'a level mismatch'.
  4. [Table 1 caption] The caption says 'Note that for fairness, Paraformer as well as Whisper use official pre-trained parameters.' This statement is ambiguous because the rows of the table are different data types (CTSS, GSS, PL), not different models; please clarify that both models are evaluated without fine-tuning on each data type.
  5. [Section 3.3 and Table 3 footnote] In the Table 3 caption, 'Session 2.2.2' should be 'Section 2.2.2'.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the pseudo-label pipeline is built from close-talk recordings and evaluated on held-out CER; the S3-versus-S1 ablation is confounded but not a circular derivation.

full rationale

The claimed derivation chain is: close-talk recordings supply high-quality speech; TLS aligns them to GSS mixtures (Eqs. 1-3); G-SpatialNet is trained with GSS and 8-channel array input against the aligned close-talk signal as target; Paraformer is fine-tuned on the enhanced audio; and held-out Dev/Eval CER is reported. Each link is measured empirically rather than defined into existence: Table 1 shows pseudo labels have CER/DNSMOS close to close-talk speech, and Table 3 reports held-out CER gains. The pseudo-label target s3(t) is not the input y(t); it is a filtered close-talk signal, so the enhancement task is not an identity mapping. Although Eq. (2) estimates the alignment filter by minimizing reconstruction error against the same y(t) that later serves as the network input, this makes the target a linear projection of y onto the close-talk subspace, not the input itself, and the final Eval CERs are independent of the training targets. The assertion that the S3 improvement of 13.5% (Dev) and 12.1% (Eval) 'fully demonstrates the effectiveness of our proposed pseudo labels-based neural speech enhancement method' is an over-attribution, because S3 differs from S1 by the entire G-SpatialNet front-end and no simulated-data or raw-close-talk training control isolates the pseudo-label contribution; this is a methodological confound, not a circular reduction. The only self-citation, [20], motivates stacking GSS and 8-channel magnitude spectrograms and is not load-bearing for the central claim. No step makes a prediction equivalent, by construction, to its input, so no circular step is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the fidelity of close-talk-derived pseudo labels and the correctness of oracle diarization, plus standard signal-processing math. No new physical entities are postulated. The free parameters are engineering hyperparameters rather than fitted scientific constants.

free parameters (6)
  • MCA loss weight alpha = not reported
    In Eq. 4, alpha weights the cosine similarity term; its value is not given although it controls the G-SpatialNet training objective.
  • MFLF filter length L = 2
    Set in Eq. 2 to L=2 T-F units; affects how well the close-talk is level-aligned to the far-field mixture.
  • SNR filter threshold = -10 dB
    Section 2.2.2 drops pseudo-label pairs below this estimated SNR; the discarded fraction is not reported.
  • Fine-tuning max epochs = 20
    Section 3.3 caps fine-tuning at 20 epochs; no search or ablation over this value is described.
  • Top-k model averaging = 5
    Section 3.3 averages the top five checkpoints by Dev CER; this is a design choice rather than an optimized parameter.
  • Weighting factor lambda(t,f) = not reported (follows [27])
    Used in Eq. 2 for MFLF; the exact weighting schedule is not given, though it shapes the fitted filter h'(f).
assumptions (5)
  • domain assumption Close-talk recordings are high-quality, leakage-minimized references for the target clean speech of each speaker.
    Used in Section 2.2.2 to justify generating pseudo labels from close-talk segments; close-talk may still contain device noise, crosstalk, or clipping, which would corrupt labels.
  • domain assumption Oracle diarization is available and correct for segmenting close-talk and far-field/GSS signals.
    Used throughout Sections 2 and 4; if diarization is wrong, both GSS and CTSS/pseudo-label pairs are misassigned, invalidating the SE training data.
  • domain assumption Magnitude spectrograms are sufficient for the ASR enhancement front-end because the Paraformer back-end uses magnitude-based features.
    G-SpatialNet discards phase in Section 2.2.1; if phase information matters for the ASR features, the enhancement may be suboptimal.
  • standard math STFT, GCC-PHAT, FCP/MFLF, and the SpatialNet architecture are correct as cited.
    The paper builds on Eqs. 1 and 2 and on SpatialNet [19] without re-deriving them; errors in these cited foundations would propagate.
  • domain assumption The GSS signals with official baseline parameters provide a suitable reference mixture for pseudo-label generation.
    TLS uses GSS signals as the reference far-field mixture in Fig. 2; if GSS leaks other speakers or removes target speech, the pseudo labels inherit those errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pseudo Labels-based Neural Speech Enhancement for the AVSR Task in the MISP-Meeting Challenge." pith.science (2026). https://pith.science/paper/WJOTNJVW

@misc{pith2026250524446,
  author       = {Pith},
  title        = {Pith review of: Pseudo Labels-based Neural Speech Enhancement for the AVSR Task in the MISP-Meeting Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJOTNJVW}},
  note         = {Machine review of arXiv:2505.24446}
}
read the original abstract

This paper presents our system for the MISP-Meeting Challenge Track 2. The primary difficulty lies in the dataset, which contains strong background noise, reverberation, overlapping speech, and diverse meeting topics. To address these issues, we (a) designed G-SpatialNet, a speech enhancement (SE) model to improve Guided Source Separation (GSS) signals; (b) proposed TLS, a framework comprising time alignment, level alignment, and signal-to-noise ratio filtering, to generate signal-level pseudo labels for real-recorded far-field audio data, thereby facilitating SE models' training; and (c) explored fine-tuning strategies, data augmentation, and multimodal information to enhance the performance of pre-trained Automatic Speech Recognition (ASR) models in meeting scenarios. Finally, our system achieved character error rates (CERs) of 5.44% and 9.52% on the Dev and Eval sets, respectively, with relative improvements of 64.8% and 52.6% over the baseline, securing second place.

Figures

Figures reproduced from arXiv: 2505.24446 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proposed TLS framework for pseudo labels estima￾tion. In this paper, the “Reference Far-Field Mixtures” are des￾ignated as GSS signals. approach, which includes the TLS framework (Section 2.2.2) for pseudo labels generation and MCA loss (2.2.3) for model training. 2.1. Guided source separation We followed the GSS implementation in [5], while using the same parameters as the official baselines1 . The GSS algorithm co… view at source ↗
Figure 4
Figure 4. Spectrogram visualization of far-field mixtures en￾hanced by GSS and our proposed “GSS+G-SpatialNet” front￾end on the MISP-Meeting evaluation set. 3.2. Fine-tuning to adapt to G-SpatialNet Since G-SpatialNet aims to minimize the MCA loss rather than the CTC/Attention loss, its enhancement direction may not nec￾essarily lead to a reduction in CER. One effective solution to this issue is to jointly train the G-Spatial… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 21 canonical work pages

  1. [1]

    Introduction With the advancement of multimodal technology, Audio-Visual Speech Recognition (A VSR) has attracted increasing attention. Previous Multimodal Information based Speech Processing (MISP) challenges released a large-scale Mandarin audio-visual conversational dataset for home-TV scenarios [1, 2, 3], promot- ing progress in far-field A VSR. The M...

  2. [2]

    Pseudo Labels-based Neural Speech Enhancement for the AVSR Task in the MISP-Meeting Challenge

    Front-end In our front-end, we first employ GSS with the oracle diariza- tion to extract each speaker’s speech signal, followed by a ded- icated SE model called G-SpatialNet (Section 2.2.1) for fur- ther signal enhancement. To enable direct training on real- world meeting data, we propose a novel pseudo labels-based arXiv:2505.24446v2 [cs.SD] 23 Jun 2025 ...

  3. [3]

    GSS+G-SpatialNet

    Back-end 3.1. ASR model We adopt Paraformer-large [13] as the back-end model, which contains 220M parameters and consists of an encoder, pre- dictor, sampler, and decoder. The continuous integrate-and- fire (CIF) [30] based predictor predicts the number of tokens and generates hidden variables. The glancing language model (GLM) [31] based sampler then gen...

  4. [4]

    Speech enhancement For G-SpatialNet training, we use only the audio data of MISP- Meeting training set

    Training data 4.1. Speech enhancement For G-SpatialNet training, we use only the audio data of MISP- Meeting training set. Specifically, We first apply GSS with the oracle diarization to pre-process the raw 8-channel far-field mi- crophone array signals. Subsequently, pseudo labels are gen- erated using the method described in Section 2.2.2. The GSS signa...

  5. [5]

    Speech enhancement: TLS and G-SpatialNet Fig

    Result and analysize 5.1. Speech enhancement: TLS and G-SpatialNet Fig. 3 illustrates an example of pseudo label estimation using our proposed TLS framework. As shown in Fig. 3, the close- talk speech fed to TLS framework time- and level-aligns with the target clean speech of the reference far-field mixtures. Fur- thermore, we evaluate the CER and speech ...

  6. [6]

    We proposed a novel framework, TLS, which gen- erates high-quality pseudo labels for real-world meeting data, enabling the direct training of SE models in real-world scenar- ios

    Conclusions In this paper, we design G-SpatialNet to improve quality of far- field audio. We proposed a novel framework, TLS, which gen- erates high-quality pseudo labels for real-world meeting data, enabling the direct training of SE models in real-world scenar- ios. Furthermore, we explore how fine-tuning strategies, data augmentation, and multimodal in...

  7. [7]

    Acknowledgments This work was supported in part by the National Natural Science Foundation of China under Grants 62371407 and 62276220, and the Innovation of Policing Science and Tech- nology, Fujian province (Grant number: 2024Y0068)

  8. [8]

    The first multimodal information based speech processing (misp) chal- lenge: Data, tasks, baselines and results,

    H. Chen, H. Zhou, J. Du, C.-H. Lee, J. Chen, S. Watanabe, S. M. Siniscalchi, O. Scharenborg, D.-Y . Liu, B.-C. Yinet al., “The first multimodal information based speech processing (misp) chal- lenge: Data, tasks, baselines and results,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 9266–9270

Show all 41 references
  1. [9]

    The mul- timodal information based speech processing (misp) 2022 chal- lenge: Audio-visual diarization and recognition,

    Z. Wang, S. Wu, H. Chen, M.-K. He, J. Du, C.-H. Lee, J. Chen, S. Watanabe, S. Siniscalchi, O. Scharenborget al., “The mul- timodal information based speech processing (misp) 2022 chal- lenge: Audio-visual diarization and recognition,” inICASSP 2023-2023 IEEE International Conf...

  2. [10]

    The multimodal information based speech processing (misp) 2023 challenge: Audio-visual target speaker extraction,

    S. Wu, C. Wang, H. Chen, Y . Dai, C. Zhang, R. Wang, H. Lan, J. Du, C.-H. Lee, J. Chenet al., “The multimodal information based speech processing (misp) 2023 challenge: Audio-visual target speaker extraction,” inICASSP 2024-2024 IEEE Interna- tional Conference on Acoustics, Sp...

  3. [11]

    Front-end processing for the chime-5 dinner party scenario,

    C. Boeddeker, J. Heitkaemper, J. Schmalenstroeer, L. Drude, J. Heymann, and R. Haeb-Umbach, “Front-end processing for the chime-5 dinner party scenario,” inCHiME5 Workshop, Hy- derabad, India, vol. 1, 2018

  4. [12]

    Gpu-accelerated guided source separation for meeting transcription,

    D. Raj, D. Povey, and S. Khudanpur, “Gpu-accelerated guided source separation for meeting transcription,” inInterspeech 2023, 2023, pp. 3507–3511

  5. [13]

    Chime-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,

    S. Watanabe, M. Mandel, J. Barker, E. Vincent, A. Arora, X. Chang, S. Khudanpur, V . Manohar, D. Povey, D. Rajet al., “Chime-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,” inCHiME 2020-6th International Workshop on Speech Processing in Every...

  6. [14]

    The chime-7 dasr challenge: Distant meeting transcrip- tion with multiple devices in diverse scenarios,

    S. Cornell, M. Wiesner, S. Watanabe, D. Raj, X. Chang, P. Gar- cia, M. Maciejewski, Y . Masuyama, Z.-Q. Wang, S. Squartini et al., “The chime-7 dasr challenge: Distant meeting transcrip- tion with multiple devices in diverse scenarios,”arXiv preprint arXiv:2306.13734, 2023

  7. [15]

    The chime- 8 dasr challenge for generalizable and array agnostic distant automatic speech recognition and diarization,

    S. Cornell, T. Park, S. Huang, C. Boeddeker, X. Chang, M. Ma- ciejewski, M. Wiesner, P. Garcia, and S. Watanabe, “The chime- 8 dasr challenge for generalizable and array agnostic distant automatic speech recognition and diarization,”arXiv preprint arXiv:2407.16447, 2024

  8. [16]

    Notsofar- 1 challenge: New datasets, baseline, and tasks for distant meeting transcription,

    A. Vinnikov, A. Ivry, A. Hurvitz, I. Abramovski, S. Koubi, I. Gur- vich, S. Peer, X. Xiao, B. M. Elizalde, N. Kandaet al., “Notsofar- 1 challenge: New datasets, baseline, and tasks for distant meeting transcription,”arXiv preprint arXiv:2401.08887, 2024

  9. [17]

    The chime- 7 udase task: Unsupervised domain adaptation for conversational speech enhancement,

    S. Leglaive, L. Borne, E. Tzinis, M. Sadeghi, M. Fraticelli, S. Wis- dom, M. Pariente, D. Pressnitzer, and J. R. Hershey, “The chime- 7 udase task: Unsupervised domain adaptation for conversational speech enhancement,” in7th International Workshop on Speech Processing in Every...

  10. [18]

    Mixture to mixture: Leveraging close-talk mixtures as weak-supervision for speech separation,

    Z.-Q. Wang, “Mixture to mixture: Leveraging close-talk mixtures as weak-supervision for speech separation,”IEEE Signal Process- ing Letters, vol. 31, pp. 1715–1719, 2024

  11. [19]

    Unssor: unsupervised neural speech separation by leveraging over-determined training mix- tures,

    Z.-Q. Wang and S. Watanabe, “Unssor: unsupervised neural speech separation by leveraging over-determined training mix- tures,”Advances in Neural Information Processing Systems, vol. 36, 2024

  12. [20]

    Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,

    Z. Gao, S. Zhang, I. McLoughlin, and Z. Yan, “Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,” inInterspeech 2022, 2022, pp. 2063– 2067

  13. [21]

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

    T. Nakatani, T. Yoshioka, K. Kinoshita, M. Miyoshi, and B.-H. Juang, “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,...

  14. [22]

    Speech dereverberation based on variance-normalized de- layed linear prediction,

    ——, “Speech dereverberation based on variance-normalized de- layed linear prediction,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, no. 7, pp. 1717–1731, 2010

  15. [23]

    Complex angular central gaus- sian mixture model for directional statistics in mask-based mi- crophone array signal processing,

    N. Ito, S. Araki, and T. Nakatani, “Complex angular central gaus- sian mixture model for directional statistics in mask-based mi- crophone array signal processing,” in2016 24th European Sig- nal Processing Conference (EUSIPCO). IEEE, 2016, pp. 1153– 1157

  16. [24]

    On optimal frequency- domain multichannel linear filtering for noise reduction,

    M. Souden, J. Benesty, and S. Affes, “On optimal frequency- domain multichannel linear filtering for noise reduction,”IEEE Transactions on audio, speech, and language processing, vol. 18, no. 2, pp. 260–276, 2009

  17. [25]

    Improved mvdr beamforming using single-channel mask prediction networks

    H. Erdogan, J. R. Hershey, S. Watanabe, M. I. Mandel, and J. Le Roux, “Improved mvdr beamforming using single-channel mask prediction networks.” inInterspeech, 2016, pp. 1981–1985

  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

  19. [27]

    The xmuspeech system for audio-visual target speaker extraction in misp 2023 challenge,

    L. Luo, T. Li, L. Li, and Q. Hong, “The xmuspeech system for audio-visual target speaker extraction in misp 2023 challenge,” in 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 2024, pp. 39– 40

  20. [28]

    Phase- sensitive and recognition-boosted speech separation using deep recurrent neural networks,

    H. Erdogan, J. R. Hershey, S. Watanabe, and J. Le Roux, “Phase- sensitive and recognition-boosted speech separation using deep recurrent neural networks,” in2015 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015, pp. 708–712

  21. [29]

    The fifth’chime’speech separation and recognition challenge: Dataset, task and baselines,

    J. Barker, S. Watanabe, E. Vincent, and J. Trmal, “The fifth’chime’speech separation and recognition challenge: Dataset, task and baselines,”Interspeech 2018, 2018

  22. [30]

    M2met: The icassp 2022 multi- channel multi-party meeting transcription challenge,

    F. Yu, S. Zhang, Y . Fu, L. Xie, S. Zheng, Z. Du, W. Huang, P. Guo, Z. Yan, B. Maet al., “M2met: The icassp 2022 multi- channel multi-party meeting transcription challenge,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). I...

  23. [31]

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

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

  24. [32]

    Robust speech recognition via large-scale weak supervision,

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

  25. [33]

    Robust localization in reverberant rooms,

    J. H. DiBiase, H. F. Silverman, and M. S. Brandstein, “Robust localization in reverberant rooms,” inMicrophone arrays: sig- nal processing techniques and applications. Springer, 2001, pp. 157–180

  26. [34]

    Convolutive predic- tion for monaural speech dereverberation and noisy-reverberant speaker separation,

    Z.-Q. Wang, G. Wichern, and J. Le Roux, “Convolutive predic- tion for monaural speech dereverberation and noisy-reverberant speaker separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3476–3490, 2021

  27. [35]

    Supervised speech separation based on deep learning: An overview,

    D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,”IEEE/ACM transactions on audio, speech, and language processing, vol. 26, no. 10, pp. 1702–1726, 2018

  28. [36]

    A simultane- ous denoising and dereverberation framework with target decou- pling,

    A. Li, W. Liu, X. Luo, G. Yu, C. Zheng, and X. Li, “A simultane- ous denoising and dereverberation framework with target decou- pling,” inInterspeech 2021, 2021, pp. 2801–2805

  29. [37]

    Cif: Continuous integrate-and-fire for end- to-end speech recognition,

    L. Dong and B. Xu, “Cif: Continuous integrate-and-fire for end- to-end speech recognition,” inICASSP 2020-2020 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 6079–6083

  30. [38]

    Glancing transformer for non-autoregressive neural machine translation,

    L. Qian, H. Zhou, Y . Bao, M. Wang, L. Qiu, W. Zhang, Y . Yu, and L. Li, “Glancing transformer for non-autoregressive neural machine translation,”arXiv preprint arXiv:2008.07905, 2020

  31. [39]

    Aishell-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,

    Y . Fu, L. Cheng, S. Lv, Y . Jv, Y . Kong, Z. Chen, Y . Hu, L. Xie, J. Wu, H. Bu, X. Xu, J. Du, and J. Chen, “Aishell-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,” inInterspeech 2021, 2021, pp. 3665–3669

  32. [40]

    Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,

    B. Zhang, H. Lv, P. Guo, Q. Shao, C. Yang, L. Xie, X. Xu, H. Bu, X. Chen, C. Zeng, D. Wu, and Z. Peng, “Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,”

  33. [2022]

    Available: https://arxiv.org/abs/2110.03370

    [Online]. Available: https://arxiv.org/abs/2110.03370

Pith tools

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