Pith. sign in

REVIEW 3 major objections 5 minor 38 references

SuPseudo: A Pseudo-supervised Learning Method for Neural Speech Enhancement in Far-field Speech Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Speech enhancement trained on pseudo-labels from close-talk microphones outperforms prior far-field ASR front-ends while using only audio.

desk verdict Solid challenge paper: DSE pseudo-labels from filtered close-talk beat simulated training and raw close-talk labels, but the fixed 5m delay assumption needs a sensitivity check. read the letter →

arxiv 2505.24450 v2 pith:P74XXDLT submitted 2025-05-30 cs.SD eess.AS

Longjie Luo , Lin Li , Qingyang Hong This is my paper · ORCID
classification cs.SDeess.AS
keywords speechenhancementpseudo-supervisedlearningfar-fieldrecognitiondirectsoundestimationdomainmismatchtargetspeakerextractionMISP2023charactererrorrate
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

Far-field speech recognition systems need speech enhancement front-ends, but enhancement models trained on simulated noisy-clean pairs degrade on real recordings because of the gap between simulated and real acoustics. This paper proposes to close that gap by estimating, for real recordings, the oracle direct sound of each target speaker from that speaker's close-talk microphone signal, and using those estimates as pseudo-labels for training. The method, called SuPseudo, lets an enhancement model learn directly from real far-field data without manual annotations or joint ASR optimization. On the MISP2023 corpus, the resulting GSS+FARNET system reaches a 29.80% character error rate, a 10.2% relative reduction over the previous state-of-the-art (33.18%), using only audio with 1.56 million parameters.

What carries the argument

The load-bearing pieces are three. Direct sound estimation (DSE) solves a multi-frame linear filtering problem per frequency, with the filter order $\hat{L}=\lceil D/(aH)\rceil+1$ coming from a fixed distance $D=5$ m, and a weighting term $\lambda(t,f)$ down-weighting low-energy T-F bins; this converts close-talk speech into an estimate of the oracle direct sound. The magnitude-constraint adjustable (MCA) loss linearly weights the Frobenius-norm MSE and a cosine-similarity term ($L_{MCA}=L_{MSE}+\alpha L_{COSSIM}$), so that the model is not forced to match misaligned pseudo-labels exactly. FARNET is an encoder-separator-decoder network that stacks the multi-channel far-field magnitudes with the GSS output as input, uses a power-law compression of $c=0.3$, four TS-Conformer blocks in the separator, and estimates a magnitude mask $M$, so that the enhanced output is $|X_q|^c=M\odot|G_q|^c$.

What would settle it

Measure the actual speaker-to-array distances in the MISP2023 training set and recompute the DSE filter order per utterance; if the resulting pseudo-labels change the evaluation CER by more than a rounding error, the fixed-distance assumption is load-bearing.

Watch

Extended reading notes

Core claim

The paper's central claim is that a real far-field recording's target direct sound can be estimated well enough to serve as a training target for speech enhancement. In the direct sound estimation (DSE) step, a per-frequency linear filter is learned by minimizing a weighted least-squares fit between the GSS-separated far-field mixture and delayed frames of the close-talk speech; the filter order comes from a fixed assumed speaker-array distance of 5 m, and the output $\hat{X}_q(t,f)=\hat{h}_q^d(f)^H \breve{Y}_q(t,f)$ is treated as the pseudo-label. Training with a magnitude-constraint adjustable (MCA) loss, which combines MSE with cosine similarity, makes the model robust to imperfect pseudo-labels. The authors report that on the MISP2023 evaluation set, the system achieves 29.80% CER, a 10.2% relative reduction over the 33.18% of the previous state of the art, while using no video and a 1.56M-parameter model.

Load-bearing premise

The weakest load-bearing assumption is that the time delay between the close-talk signal and the far-field direct sound is well approximated by a fixed speaker-array distance of 5 m in Eq. (5).

Editorial extensions

If this is right

  • Far-field enhancement models can be trained or fine-tuned on real conversational data wherever a close-talk reference exists, removing the need for clean-speech annotations.
  • Signal-level pseudo-supervision transfers across ASR back-ends, whereas joint optimization with one ASR model degraded performance on other decoders (Paraformer, Whisper).
  • Two-stage training, simulated pre-training followed by SuPseudo fine-tuning, outperforms either simulated-only or real-only training.
  • A small audio-only front-end using this training can beat multi-system, audio-visual fusions on the MISP2023 benchmark.

Reading between the lines

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

  • If DSE generalizes, any meeting scenario with lapel or close-talk microphones and a far-field array could reuse this pseudo-label recipe, not just the MISP setup.
  • The MCA loss is a generic label-noise tolerance trick; it could be tested on other pseudo-label or self-training pipelines in speech and vision.
  • Replacing GSS with a learned target-speaker-extraction front-end and keeping SuPseudo might yield a fully data-driven distant-conversation front-end, which the authors state as future work.
  • The fixed-distance filter-order formula may be a limitation that an adaptive distance estimate would remove; this is not explored in the paper.
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 / 5 minor

Summary. The paper proposes a pseudo-supervised learning method, SuPseudo, for speech enhancement in far-field speech recognition. It introduces direct sound estimation (DSE), a signal-processing pipeline that estimates the oracle direct sound of real-recorded data from close-talk recordings and guided source separation (GSS) outputs, and uses these estimates as pseudo-labels to train a compact neural enhancement model called FARNET. A magnitude constraint adjustable (MCA) loss is designed to tolerate imperfections in the pseudo-labels. Experiments on the MISP2023 corpus show that the proposed system achieves 29.80% character error rate (CER) on the evaluation set, outperforming the previous state-of-the-art single-system and multi-system fusion results, and ablations in Table 1 demonstrate the contribution of each component.

Significance. If the reported results are reliable, this work is significant because it provides a practical way to adapt speech enhancement models to real-recorded far-field data without manual target-speech annotations, addressing a long-standing domain-mismatch problem. The empirical core is well supported: Table 1 shows that DSE pseudo-labels outperform close-talk segments and simulated-data-only training, and Table 2 shows consistent gains across three different ASR back-ends while using only audio and 1.56M parameters. The paper also contributes a careful comparison against top MISP2023 challenge systems. However, the central novelty DSE lacks direct validation against ground-truth direct sound, and its derivation depends on approximations that deserve scrutiny.

major comments (3)
  1. [Section 3.2, Eq. (7)] The derivation of the DSE optimization problem relies on W-disjoint orthogonality to drop the cross-term between (R_q - \hat{h}^H \breve{S}_q) and I_q, but (R_q - \hat{h}^H \breve{S}_q) is not a source signal and may not be disjoint from I_q in the time-frequency domain; this step is therefore not justified by the cited assumption. Moreover, the claim that the early-reflection term h^{e,l}_q^H \breve{S}_q(t,f) does not affect the filter estimate presupposes that the filter order \hat{L} exactly matches the direct-sound delay, which is not demonstrated. Since the paper never validates DSE outputs against ground-truth direct sound (e.g., on simulated data where the true direct sound is known), the central mechanism is unverified. Please add an experiment on simulated data reporting the accuracy of DSE estimates (e.g., SI-SDR or PESQ against the known direct sound) or a quantitative analysis of the residual error, to substantiate the claim that DSE estimates the oracle direct sound.
  2. [Section 3.2, Eq. (5)] The filter order \hat{L} is computed from a fixed speaker-array distance D = 5 m, but in the MISP2023 recording scenario speakers sit at and move between varying distances. A misestimated \hat{L} will misalign the pseudo-labels, and while the MCA loss is intended to mitigate label errors, cosine similarity is not invariant to a multi-frame time shift and cannot fully correct a systematic delay. The authors should provide a sensitivity analysis on the development set for a range of D values (e.g., 1, 3, 5, and 8 m) and discuss whether the method can be applied to datasets without a known typical distance. Without such an analysis, the portability of the method beyond MISP2023 is not established.
  3. [Section 4, data preparation] The method uses oracle diarization for both GSS separation and close-talk segmentation. This is available in the MISP2023 challenge setup but is a strong assumption for real-world deployment. The authors should explicitly discuss this limitation and, if feasible, report an experiment with an estimated diarization system to quantify the impact on the final CER.
minor comments (5)
  1. [Conclusion] There is a typo in the conclusion: "leaveage" should be "leverage".
  2. [Section 3.2, Eq. (7)] The derivation in Eq. (7) is hard to follow because the optimization objective is written without an explicit argmin variable in some rows; please clarify the notation for readability.
  3. [Table 1 note] The sentence "Systems B1 refers to 'GSS', while systems M2-M4d refer to 'GSS+FARNET'" is grammatically inconsistent; rephrase to "System B1 refers to 'GSS', while systems M2-M4d refer to 'GSS+FARNET'".
  4. [Section 1, Introduction] The term "pseudo-supervised learning" is cited to [18], a paper on deep subspace clustering; a more standard reference on self-training or pseudo-labeling would better situate the method.
  5. [Table 2] The claim of "significantly outperforms" is not supported by statistical significance tests; consider softening the wording to "outperforms" to match the reporting.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: DSE pseudo-labels are produced by a fixed signal-processing pipeline from an independent close-talk reference, and the reported gains are measured on a held-out evaluation set with external ASR models.

full rationale

The paper's central derivation chain is not circular. DSE estimates the oracle direct sound by solving a fixed least-squares filtering problem (Eq. 4) that maps delayed close-talk segments to the GSS output; the resulting pseudo-labels are computed once by a signal-processing procedure and are independent of the FARNET model parameters. SuPseudo then trains FARNET on these fixed pseudo-labels, which is a standard supervised/self-training setup rather than a derivation that assumes its own conclusion. The ablation in Table 1 (M3 vs. M2, M4c vs. M4d) and the comparison in Table 2 are evaluated on the held-out MISP2023 evaluation set using the official ASR model plus two external universal ASR models (Paraformer-large and Whisper-large-v3) and DNSMOS, so the reported CER improvements are not forced by construction. The only self-citation is [22], the authors' own MISP2023 challenge system, which is used both as a baseline (S1) and as the source of the input-stacking design choice; this citation is not load-bearing for the claimed result and does not constitute circularity. Concerns about the fixed distance D = 5 m in Eq. (5) and the W-disjoint orthogonality assumption in Eq. (7) are correctness or robustness risks about pseudo-label quality, not circularity: an inaccurate or misaligned label would degrade training, but the method does not define the label to be identical to the model output. The derivation therefore stands as an independent estimation-plus-training pipeline, and the score reflects only the presence of a minor, non-load-bearing self-citation.

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

The central claim rests on the DSE pseudo-label quality, which depends on assumptions about close-talk fidelity, W-disjoint orthogonality, and a fixed speaker-array distance. The free parameters (D, \hat{L}, \epsilon, \alpha, c) are all set by hand or from prior work, with no sensitivity analysis. No new physical entities are introduced.

free parameters (5)
  • Distance D between speaker and array = 5 m
    Used in Eq. (5) to compute the DSE filter delay; assumed known and fixed, but actual speaker distances may vary.
  • DSE filter order \hat{L} = 4
    Derived from D=5 m in Eq. (5); effectively a hand-set hyperparameter because D is fixed.
  • Weight flooring factor \epsilon = 0.01
    Appears in the DSE weighting term Eq. (6); no sensitivity analysis is given.
  • MCA loss weight \alpha = 0.2
    Controls trade-off between MSE and cosine-similarity terms in Eq. (8); no sensitivity analysis is given.
  • Power-law compression exponent c = 0.3
    Adopted from Braun and Tashev [21] for the magnitude spectrogram input.
assumptions (5)
  • domain assumption Close-talk microphone segment approximates the dry source signal S_q(t,f) (Eq. 3).
    The paper assumes the close-talk recording is nearly clean and well-aligned to the source; it acknowledges possible leakage but does not model it.
  • domain assumption W-disjoint orthogonality of target speech and residual interference in the STFT domain.
    Used in the DSE derivation (Eq. 7) to drop cross terms; real speech mixtures only approximately satisfy this.
  • domain assumption Narrowband convolutive transfer function approximation for direct and non-direct sound paths.
    Eq. (2) models direct and non-direct speech as short linear filters on delayed copies of the dry source; standard but approximate.
  • domain assumption GSS with oracle diarization produces G_q(t,f) = R_q(t,f) + I_q(t,f) with manageable residual interference.
    The entire pipeline rests on GSS preprocessing; if GSS fails, both the pseudo-labels and the model input degrade.
  • domain assumption Fixed and known speaker-array distance D=5 m for all conversations.
    Eq. (5) sets the filter order from D; no variability or uncertainty is considered.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SuPseudo: A Pseudo-supervised Learning Method for Neural Speech Enhancement in Far-field Speech Recognition." pith.science (2026). https://pith.science/paper/P74XXDLT

@misc{pith2026250524450,
  author       = {Pith},
  title        = {Pith review of: SuPseudo: A Pseudo-supervised Learning Method for Neural Speech Enhancement in Far-field Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P74XXDLT}},
  note         = {Machine review of arXiv:2505.24450}
}
read the original abstract

Due to the lack of target speech annotations in real-recorded far-field conversational datasets, speech enhancement (SE) models are typically trained on simulated data. However, the trained models often perform poorly in real-world conditions, hindering their application in far-field speech recognition. To address the issue, we (a) propose direct sound estimation (DSE) to estimate the oracle direct sound of real-recorded data for SE; and (b) present a novel pseudo-supervised learning method, SuPseudo, which leverages DSE-estimates as pseudo-labels and enables SE models to directly learn from and adapt to real-recorded data, thereby improving their generalization capability. Furthermore, an SE model called FARNET is designed to fully utilize SuPseudo. Experiments on the MISP2023 corpus demonstrate the effectiveness of SuPseudo, and our system significantly outperforms the previous state-of-the-art. A demo of our method can be found at https://EeLLJ.github.io/SuPseudo/.

Figures

Figures reproduced from arXiv: 2505.24450 by the authors.

Figure 1
Figure 1. Schematic diagram of the recording scenario. data by leveraging the weak supervision of word transcrip￾tions. Additionally, the perceptual evaluation of speech quality (PESQ) estimators [15] and the DNSMOS models [16] can also provide reference-free losses for fine-tuning SE models on real￾recorded data. Recently, inspired by the blind deconvolution problem, Wang [17] designed some loss functions that enforce the mo… view at source ↗
Figure 2
Figure 2. Illustration of the proposed method. In SuPseudo, if the model has been pre-trained on simulated data, then the fine￾tuning (ft.) is performed; otherwise, the training (trng.) directly on real-recorded data is performed. This paper aims to further reconstruct the target speaker q’s direct sound h d q (f) HS¯q(t, f) from the GSS-output Gq(t, f), so as to improve the speech quality and the ASR performance. 3. Proposed… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    Introduction Speech enhancement (SE) algorithms can significantly improve automatic speech recognition (ASR) performance in far-field conversational scenarios [1]. Meanwhile, modern data-driven SE models have showcased powerful performance [2, 3, 4], at- tracting increasing researchers to explore their application in real-world far-field speech recognitio...

    work page Pith review arXiv 2022

  2. [2]

    1,Nspeakers are spontaneously grouped for daily conversation in a reverberant room

    Problem Formulation As shown in Fig. 1,Nspeakers are spontaneously grouped for daily conversation in a reverberant room. Each speaker wears a high-fidelity microphone (i.e., close-talk microphone) posi- tioned at the center of their chin. Additionally, aP-channel mi- crophone array (i.e., far-field microphone array) is placed near a TV . Other background ...

    work page

  3. [3]

    GSS”, while systems M2-M4d refer to “GSS+FARNET

    Proposed Method Fig. 2 illustrates the proposed SuPseudo training and FARNET model. As described in Section 2, we first use GSS [19], along with the oracle diarization, to pre-processP-channel far-field recordings. The proposed FARNET model is then used to fur- ther enhance the GSS-outputG q(t, f). For SuPseudo training, the DSE-estimates are used as pseu...

    work page

  4. [4]

    GSS+MEASE

    Experimental Setup For dataset, we utilize the MISP2023 corpus [5], a far-field multi-speaker conversational Chinese audio-visual corpus, to evaluate our method. This corpus focuses on real home-TV scenarios (as shown in Fig. 1), where 2-6 speakers (i.e.,N∈ {2, ...,6}) spontaneously communicate without specific topics under strong background noise and rev...

    work page

  5. [5]

    GSS+FARNET

    Results And Analysis In Table 1, the impact of various training strategies on model performance is illustrated. It is clear that systems M2-M4d all outperform GSS. This suggests that the GSS-output still con- tains significant interferences and highlights the effectiveness of the FARNET model. While system M2, which employs a su- pervised learning method ...

    work page

  6. [6]

    GSS+FARNET

    Conclusion In this paper, we propose DSE to estimate the oracle direct sound of real-recorded data for SE. And we present SuPseudo, a novel training framework to leaveage the DSE-estimates and enhance the models’ generalization capability to real-world conditions. To better match far-field speech recognition, we adopt the FARNET model as the backbone. Abl...

    work page

  7. [7]

    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

    work page 2018

  8. [8]

    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

    work page 2018

Show all 38 references
  1. [9]

    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,” inICASSP 2021-2021 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6633–6637

    2021

  2. [10]

    Efficient monaural speech enhancement with universal sample rate band-split rnn,

    J. Yu and Y . Luo, “Efficient monaural speech enhancement with universal sample rate band-split rnn,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  3. [11]

    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

  4. [12]

    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...

    2023

  5. [13]

    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

    2024 arXiv

  6. [14]

    An audio-quality-based multi-strategy approach for target speaker extraction in the misp 2023 challenge,

    R. Han, X. Yan, W. Xu, P. Guo, J. Sun, H. Wang, Q. Lu, N. Jiang, and L. Xie, “An audio-quality-based multi-strategy approach for target speaker extraction in the misp 2023 challenge,” in2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (I...

    2023

  7. [15]

    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...

    2020

  8. [16]

    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

    2023 arXiv

  9. [17]

    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 Co...

    2022

  10. [18]

    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...

    2022

  11. [19]

    Ef- fects of lombard reflex on the performance of deep-learning-based audio-visual speech enhancement systems,

    D. Michelsanti, Z.-H. Tan, S. Sigurdsson, and J. Jensen, “Ef- fects of lombard reflex on the performance of deep-learning-based audio-visual speech enhancement systems,” inICASSP 2019- 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE...

    2019

  12. [20]

    of target direct sound, withh d q (f)∈C L being a linear filter and ¯Sq(t, f) = [Sq(t, f), ..., Sq(t−L+ 1, f)] T stack- ingLdelayed copies of the dry source signalS q(t, f). Simi- larly,h e,l q (f) H eSq(t, f)denotes the narrowband approximation of target non-direct sound (i.e...

  13. [21]

    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...

    2023

  14. [22]

    Employing real training data for deep noise suppression,

    Z. Xu, M. Sach, J. Pirklbauer, and T. Fingscheidt, “Employing real training data for deep noise suppression,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 731–10 735

    2024

  15. [23]

    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

    2022

  16. [24]

    Superm2m: Supervised and mixture-to-mixture co-learning for speech enhancement and robust asr,

    Z.-Q. Wang, “Superm2m: Supervised and mixture-to-mixture co-learning for speech enhancement and robust asr,” 2024. [Online]. Available: https://arxiv.org/abs/2403.10271

    2024 arXiv

  17. [25]

    Pseudo-supervised deep subspace clustering,

    J. Lv, Z. Kang, X. Lu, and Z. Xu, “Pseudo-supervised deep subspace clustering,”IEEE Transactions on Image Processing, vol. 30, pp. 5252–5263, 2021

    2021

  18. [26]

    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

    2023

  19. [27]

    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

  20. [28]

    A consolidated view of loss functions for supervised deep learning-based speech enhancement,

    S. Braun and I. Tashev, “A consolidated view of loss functions for supervised deep learning-based speech enhancement,” in2021 44th International Conference on Telecommunications and Signal Processing (TSP). IEEE, 2021, pp. 72–76

    2021

  21. [29]

    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

    2023

  22. [30]

    A deep ensemble learning method for monaural speech separation,

    X.-L. Zhang and D. Wang, “A deep ensemble learning method for monaural speech separation,”IEEE/ACM transactions on audio, speech, and language processing, vol. 24, no. 5, pp. 967–977, 2016

    2016

  23. [31]

    Densely connected neural network with dilated convolutions for real-time speech enhancement in the time domain,

    A. Pandey and D. Wang, “Densely connected neural network with dilated convolutions for real-time speech enhancement in the time domain,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 6629–6633

    2020

  24. [32]

    Instance normaliza- tion: The missing ingredient for fast stylization,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Instance normaliza- tion: The missing ingredient for fast stylization,”arXiv preprint arXiv:1607.08022, 2016

    2016 arXiv

  25. [33]

    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

    2021

  26. [34]

    Blind separation of speech mixtures via time-frequency masking,

    O. Yilmaz and S. Rickard, “Blind separation of speech mixtures via time-frequency masking,”IEEE Transactions on signal pro- cessing, vol. 52, no. 7, pp. 1830–1847, 2004

    2004

  27. [35]

    Improving audio-visual speech recognition by lip-subword correlation based visual pre-training and cross-modal fusion en- coder,

    Y . Dai, H. Chen, J. Du, X. Ding, N. Ding, F. Jiang, and C.-H. Lee, “Improving audio-visual speech recognition by lip-subword correlation based visual pre-training and cross-modal fusion en- coder,” in2023 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2023...

    2023

  28. [36]

    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, ser. Interspeech, 2022, pp. 2063–2067, interspeech Conference, Incheon, SOUTH KOREA, SEP 18-22, 2022

    2022

  29. [37]

    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

    2023

  30. [38]

    Sir- progressive audio-visual tf-gridnet with asr-aware selector for tar- get speaker extraction in misp 2023 challenge,

    Z. Hou, T. Sun, Y . Hu, C. Zhu, K. Chen, and J. Lu, “Sir- progressive audio-visual tf-gridnet with asr-aware selector for tar- get speaker extraction in misp 2023 challenge,” in2024 IEEE In- ternational Conference on Acoustics, Speech, and Signal Process- ing Workshops (ICASSP...

    2023

Pith tools

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