Pith. sign in

REVIEW 3 major objections 5 minor 38 references

A Lightweight Hybrid Dual Channel Speech Enhancement System under Low-SNR Conditions

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

Pith's one-line read A 24k-parameter dual-channel network fed with blind-separation estimates outperforms larger baselines at very low SNR.

desk verdict A credible lightweight dual-channel enhancement recipe that stacks a known BSS front-end under a dual-channel GTCRN; the headline gains are plausible but hang on an underspecified source-labeling step and single-run results. read the letter →

arxiv 2505.19597 v1 pith:AS24KS3P submitted 2025-05-26 eess.AS

classification eess.AS
keywords speechenhancementdual-channelindependentvectoranalysisAux-IVAGTCRNlow-SNRcomplexratiomasklightweightmodel
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 tries to establish that a traditional blind source separation algorithm can serve as a cheap front-end to a lightweight deep-learning speech enhancer, making the combined system beat larger neural baselines in very noisy two-microphone settings. The proposed system passes the separated speech and noise from Aux-IVA, together with the original noisy spectrogram, into a modified grouped temporal convolutional recurrent network. On simulated dual-channel data, the full system is reported to lead all baselines on perceptual quality, intelligibility, and non-intrusive quality scores at -12.5 dB, -7.5 dB, and -2.5 dB SNR, while keeping only 24.39k parameters and 43.20 MMACs/s of computation. If that result is right, it gives a concrete recipe for real-time low-SNR enhancement on resource-constrained devices.

What carries the argument

The key machinery is the Aux-IVA front-end feeding auxiliary source estimates into a modified GTCRN. Aux-IVA is a fast implementation of independent vector analysis that iteratively updates demixing matrices to produce coarse separated speech and noise spectrograms. The network receives both the original noisy complex spectrogram and the selected IVA features concatenated along the channel dimension; the best configuration uses log-power spectrograms of both separated sources. Inside the network, a band merging module maps high-frequency bands onto ERB bands, a grouped dual-path RNN models intra- and inter-frame structure, and the decoder outputs a complex ratio mask that is applied to the original noisy input. A dual-encoder variant was also tested but abandoned because the lightweight network could not exploit it.

What would settle it

Compute, frame by frame, whether the Aux-IVA output labeled 'speech' is closer to the true clean speech than the output labeled 'noise' on the paper's simulated test set; if the labels swap or the relative scale changes across frames, the auxiliary input is not stable and the reported gains would not be expected to transfer to real recordings.

Watch

Extended reading notes

Core claim

The central discovery is that combining Aux-IVA as a coarse estimator with a modified dual-channel GTCRN yields the best enhancement scores among all compared systems on the simulated test set. The separated speech and noise from IVA are used as auxiliary input features, and the best configuration uses log-power spectrograms of both separated sources while applying the predicted complex ratio mask to the original noisy input rather than to the IVA estimate. At the hardest tested SNR of -12.5 dB, the proposed system reaches PESQ 1.39, against 1.15 for the dual-channel GTCRN baseline and 1.17 for the larger GTCRN-L baseline, and it likewise leads on STOI and DNSMOS at every tested SNR. This is achieved with 24.39k parameters and 43.20 MMACs/s, where the Aux-IVA module contributes only 0.20 MMACs/s per iteration. The system's claim is therefore that substantial low-SNR enhancement gains are available at nearly the same parameter and computation cost as the dual-channel baseline.

Load-bearing premise

The load-bearing premise is that the two outputs of the blind separation step keep a consistent channel order and scale from frame to frame, so the network always receives the speech estimate and the noise estimate in the same slot.

Editorial extensions

If this is right

  • The system is small enough for real-time edge deployment: 24.39k parameters and 43.20 MMACs/s, with Aux-IVA contributing only 0.20 MMACs/s per iteration.
  • The larger neural baseline DC-GTCRN-L, at 34.80k parameters, still trails the proposed system on every reported metric and SNR, so the gain does not come from raw model size.
  • Both separated sources are useful: including the IVA noise estimate alongside the speech estimate improves most metrics over using the speech estimate alone.
  • Applying the predicted mask to the original noisy mixture works better than applying it to the IVA estimate, so the network is best used as a refiner of the mixture rather than of the separation output.
  • The tested dual-encoder input path hurt rather than helped, indicating that adding input-path capacity is not a reliable way to exploit auxiliary information in a lightweight model.

Reading between the lines

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

  • The paper does not test what happens when the two IVA channels swap order between frames; swapping them before concatenation would show whether the network has learned to rely on a stable permutation.
  • A testable extension is to substitute other cheap source estimates, such as beamformer outputs or spectral-subtraction masks, for Aux-IVA to see whether the benefit is tied to IVA's harmonic-structure preservation or to any structured auxiliary input.
  • The paper's experiments are limited to simulated room impulse responses with RT60 between 0.1 and 0.4 s and a fixed two-microphone geometry, so the claimed margins could shrink under stronger reverberation, speaker movement, or different array spacing.
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. This paper proposes a hybrid dual-channel speech enhancement system that combines Aux-IVA as a blind source separation front-end with a modified dual-channel GTCRN. The IVA outputs (speech and noise estimates) are concatenated with the noisy spectrogram as auxiliary inputs. The authors ablate feature types (complex vs LPS), auxiliary channels (speech only vs speech+noise), masking strategies, and encoder count, and compare against GTCRN, DC-GTCRN, DC-GTCRN-L on simulated two-microphone low-SNR data. The proposed configuration (ID-6) attains the best PESQ/STOI/DNSMOS scores in Table 2, e.g., PESQ 1.39 vs 1.15 for DC-GTCRN at -12.5 dB, with 24.39k parameters and 43.20 MMACs/s.

Significance. The result, if reproducible, is a useful practical contribution: it improves low-SNR two-channel enhancement by a wide margin over the dual-channel baseline at nearly identical parameter count, and the ablation logic supports the design choices. The paper also releases audio examples for subjective inspection. However, the current manuscript leaves unspecified a load-bearing detail—how the two IVA outputs are labeled as speech versus noise—and reports no estimate of variability; both issues need to be resolved before the central claim can be fully accepted.

major comments (3)
  1. [3.1, 3.2, and Table 2] The manuscript never states how the two outputs of Aux-IVA are assigned to the 'speech' and 'noise' auxiliary channels. Equations (5)-(8) define the demixing-matrix update rules but impose no canonical source order; initializing W(k) to the identity matrix does not guarantee that output 1 remains the speech source after convergence. Because the central result in Table 2 (and the ID-6 vs ID-5 comparison in Table 1) depends on concatenating the two separated channels as consistently labeled 'S&N' features, the system is undefined unless an alignment rule is given. The rule must be executable at inference time without access to the clean signal; if the assignment is instead resolved by matching to the clean reference during training, that is oracle information and must be disclosed. Without this specification, the reported gains may not transfer to real recordings.
  2. [Tables 1 and 2] All reported results are point estimates from a single training run and a single test evaluation. There are no error bars, confidence intervals, or significance tests. While the headline PESQ gap at -12.5 dB (1.39 vs 1.15 for DC-GTCRN) is large, some DNSMOS differences are small (e.g., DNSMOS-BAK 3.76 vs 3.60 at -12.5 dB), and without variability estimates the reader cannot judge whether the improvements are stable across training seeds or consistent across the 500-utterance test set. Please report mean and standard deviation over at least three training runs, or provide utterance-level confidence intervals for the test-set metrics.
  3. [Table 2, Aux-IVA row] The 'Aux-IVA' baseline row also presupposes a specific way to choose which of the two separated outputs is the speech estimate, but the selection rule is not described. If the output is chosen as the one with higher correlation to the clean speech reference, then this baseline is an oracle upper bound, which would make the comparison to the proposed method more favorable than is fair. Please state how the Aux-IVA row was computed, and if an oracle selection was used, replace it with an inference-time rule (e.g., fixed output index with the same alignment used for the proposed method, or a permutation-invariant evaluation).
minor comments (5)
  1. [Table 2] The headers 'DNSMOS-SIG', 'DNSMOS-BAK', 'DNSMOS-OVRL' are unexplained; Table 1 labels these subcolumns as DNSMOS-P.835. Please use consistent notation.
  2. [Section 4.2] The cost of Aux-IVA is given as '0.20 MMACs per second per iteration'; since IVA is iterative, report the total cost per audio second for the actual number of iterations used, or clarify the statement.
  3. [Section 5.3] The conclusions are drawn only from simulated data; a short discussion of expected behavior on real recordings (or an additional real-recorded test set) would strengthen the practical-claim framing.
  4. [Equation (9)] The roles of the two weighting factors alpha and beta are not discussed beyond their values; a sentence explaining why alpha=0.01 and beta=0.3 were chosen would help.
  5. [Figures 1 and 3] The spectrogram figures would benefit from stating the SNR and the STFT parameters used for visualization; currently the reader cannot relate them to the experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed gains are measured on a held-out simulated test set and are not forced by fitted parameters, self-citations, or by-construction identities.

full rationale

The paper makes an empirical claim: that adding Aux-IVA auxiliary features to a modified dual-channel GTCRN improves PESQ, STOI, and DNSMOS on a held-out simulated test set (Table 2). This is a measured performance comparison, not a derivation whose output is equivalent to its inputs. The IVA update rules (Eqs. 5-8) are standard auxiliary-function IVA, and the GTCRN architecture is prior published work used as a baseline; citing one's own prior architecture is not circular because the proposal is evaluated against that architecture rather than derived from it. The reported parameter counts and MACs are stated for all models, and the test SNR levels (-12.5, -7.5, -2.5 dB) are distinct from the training SNR range (-10 to 0 dB), so the comparison is not statistically forced. The skeptic's concern about permutation/scaling ambiguity of IVA outputs is a missing specification about whether the speech/noise auxiliary channels are stably labeled across frames; that is an external-validity or implementation-correctness risk, not a circularity of the paper's argument. No fitted parameter is renamed as a prediction, and no load-bearing step reduces by construction to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on conventional simulated data, a standard BSS algorithm, a prior lightweight DNN, and hand-chosen loss weights. The main manual choices are the loss weights, the architecture hyperparameters, and the untested assumption that IVA channel order is stable. There are no new physical entities or fitted constants of nature; the paper is an empirical system study rather than a derivation.

free parameters (3)
  • loss weight alpha = 0.01
    Weight on the SISNR term in Eq. (9); chosen manually and not justified or optimized, it influences the training balance and final metric scores.
  • loss weight beta = 0.3
    Weight balancing compressed magnitude and real/imaginary MSE terms in Eq. (9); chosen manually and affects convergence quality and the reported results.
  • architecture hyperparameters = encoder channels 16/12, GT-Conv dilation 1/2/5, kernel sizes
    Hand-designed capacity and dilation schedule in Section 4.2 that sets the model's complexity and performance; no search or sensitivity analysis is reported. These are design choices the result depends on, not learned weights.
assumptions (4)
  • standard math Aux-IVA update rules (Eqs. 6-8) converge to valid demixing matrices for the chosen contrast function G(·).
    Section 3.2 imports the algorithm from Ono (2011) and Taniguchi et al. (2014); no proof or iteration count is given.
  • domain assumption Image-method room impulse responses with RT60 0.1-0.4 s and source-array distances 0.5-3 m are representative of real low-SNR dual-channel speech enhancement conditions.
    Section 4.1 defines the simulated dataset; the central performance claim is measured only on this simulated test set, so real-world applicability depends on this assumption.
  • domain assumption Aux-IVA outputs keep a stable 'speech' and 'noise' channel order that can be used directly as auxiliary features without permutation or scaling alignment.
    Section 3.1 concatenates IVA outputs as speech and noise without any alignment mechanism; IVA's known source-order ambiguity is not discussed, and the benefit relies on consistent ordering.
  • domain assumption The hybrid loss in Eq. (9) with the stated weights is a valid proxy for perceptual quality and intelligibility in low-SNR conditions.
    Section 3.5 combines SISNR and compressed MSE terms; the paper does not provide evidence that this loss is better than alternatives for this task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Lightweight Hybrid Dual Channel Speech Enhancement System under Low-SNR Conditions." pith.science (2026). https://pith.science/paper/AS24KS3P

@misc{pith2026250519597,
  author       = {Pith},
  title        = {Pith review of: A Lightweight Hybrid Dual Channel Speech Enhancement System under Low-SNR Conditions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AS24KS3P}},
  note         = {Machine review of arXiv:2505.19597}
}
read the original abstract

Although deep learning based multi-channel speech enhancement has achieved significant advancements, its practical deployment is often limited by constrained computational resources, particularly in low signal-to-noise ratio (SNR) conditions. In this paper, we propose a lightweight hybrid dual-channel speech enhancement system that combines independent vector analysis (IVA) with a modified version of the dual-channel grouped temporal convolutional recurrent network (GTCRN). IVA functions as a coarse estimator, providing auxiliary information for both speech and noise, while the modified GTCRN further refines the speech quality. We investigate several modifications to ensure the comprehensive utilization of both original and auxiliary information. Experimental results demonstrate the effectiveness of the proposed system, achieving enhanced speech with minimal parameters and low computational complexity.

Figures

Figures reproduced from arXiv: 2505.19597 by the authors.

Figure 1
Figure 1. An example of IVA separation performance under low SNR conditions: left shows the dual-channel noisy spec￾trograms, and right shows the separated signal spectrograms. In response to the need for practical applications, recent works have focused on developing lightweight models that maintain competitive performance with reduced computational resources. For example, RNNoise [9] first performs at a low resolution and t… view at source ↗
Figure 2
Figure 2. (a) The framework of our proposed system, (b) two types of the feature selection module, where “S or S&N” refers to using either the speech channel alone or two separated channels, Re(·) and Im(·) represent the real and imaginary parts of the complex spectrogram, respectively, and LPS stands for log-power spectrogram, (c) the encoder module input features, auxiliary information, and masking approaches, along with th… view at source ↗
Figure 3
Figure 3. Typical spectrograms A set of typical audio samples is presented in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] 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]

    A Lightweight Hybrid Dual Channel Speech Enhancement System under Low-SNR Conditions

    Introduction Speech enhancement aims to extract desired speech signals de- graded by noise and interference. It serves as a crucial front- end module in applications, including human-machine inter- action, video conferencing [1], and hearing aids [2, 3]. With the rapid advancement of deep neural networks (DNNs), data- driven speech enhancement systems hav...

  2. [2]

    Problem formulation We consider the task of extracting the target speech signal from ambient noise and other interference. The problem in the time domain can be expressed as follows: xm(t) =hm(t) ∗ s(t), (1) ym(t) =xm(t) +nm(t), (2) where t is the time index, s(t) represents the non-reverberant speech signal, hm(t) represents the room impulse response (RI...

  3. [3]

    We introduce the dual-channel version of the model and incorporate several modifications to ensure the comprehensive utilization of diverse information

    Proposed system As depicted in Figure 2a, our proposed system is built upon the GTCRN architecture. We introduce the dual-channel version of the model and incorporate several modifications to ensure the comprehensive utilization of diverse information. These modi- fications include feature selection, auxiliary information selec- tion, masking approach sel...

  4. [4]

    Dataset We generate the simulated dataset with the image method [23], with dual-channel RIRs based on a linear array with two mi- crophones placed 4 cm apart

    Experiment 4.1. Dataset We generate the simulated dataset with the image method [23], with dual-channel RIRs based on a linear array with two mi- crophones placed 4 cm apart. The room size ranges from 3 m×3 m×2.5 m to 10 m ×10 m×3 m, and the reverberation time (RT60) ranges from 0.1 s to 0.4 s. The distance from the source to the array is randomly selecte...

  5. [5]

    Results 5.1. Evaluation metrics The evaluation is conducted using the objective metrics, includ- ing perceptual evaluation of speech quality (PESQ) [27] and short-time objective intelligibility (STOI) [28]. Additionally, DNN-based non-intrusive subjective metrics DNSMOS P.808

  6. [6]

    Aux-IV A acts as a coarse esti- mator, providing auxiliary information, while the GTCRN fur- ther refines the speech quality

    Conclusion In this paper, we propose a hybrid dual-channel speech en- hancement system designed for low-SNR conditions, integrat- ing IV A and a modified GTCRN. Aux-IV A acts as a coarse esti- mator, providing auxiliary information, while the GTCRN fur- ther refines the speech quality. Through various architecture modifications, both the original and auxi...

  7. [7]

    12274221) and the AI & AI for Science Project of Nanjing University

    Acknowledgements This work is supported by the National Natural Science Foun- dation of China (Grant No. 12274221) and the AI & AI for Science Project of Nanjing University

  8. [8]

    Adl- mvdr: All deep learning mvdr beamformer for target speech sepa- ration,

    Z. Zhang, Y . Xu, M. Yu, S.-X. Zhang, L. Chen, and D. Yu, “Adl- mvdr: All deep learning mvdr beamformer for target speech sepa- ration,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6089–6093

Show all 38 references
  1. [9]

    Real-time speech enhance- ment using an efficient convolutional recurrent network for dual- microphone mobile phones in close-talk scenarios,

    K. Tan, X. Zhang, and D. Wang, “Real-time speech enhance- ment using an efficient convolutional recurrent network for dual- microphone mobile phones in close-talk scenarios,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)....

  2. [10]

    A smart binaural hearing aid architecture leveraging a smartphone app with deep- learning speech enhancement,

    Y . Li, F. Chen, Z. Sun, J. Ji, W. Jia, and Z. Wang, “A smart binaural hearing aid architecture leveraging a smartphone app with deep- learning speech enhancement,” IEEE Access, vol. 8, pp. 56 798– 56 810, 2020

  3. [11]

    Design of matrix wiener filter for noise reduction and speech enhancement in hearing aids,

    N. Modhave, Y . Karuna, and S. Tonde, “Design of matrix wiener filter for noise reduction and speech enhancement in hearing aids,” in 2016 IEEE International Conference on Recent Trends in Elec- tronics, Information & Communication Technology (RTEICT) . IEEE, 2016, pp. 843–847

  4. [12]

    Unetgan: A robust speech enhancement approach in time domain for ex- tremely low signal-to-noise ratio condition,

    X. Hao, X. Su, Z. Wang, H. Zhang, and Batushiren, “Unetgan: A robust speech enhancement approach in time domain for ex- tremely low signal-to-noise ratio condition,” in Interspeech 2019, 2019, pp. 1786–1790

  5. [13]

    Masking and inpainting: A two-stage speech enhancement ap- proach for low snr and non-stationary noise,

    X. Hao, X. Su, S. Wen, Z. Wang, Y . Pan, F. Bao, and W. Chen, “Masking and inpainting: A two-stage speech enhancement ap- proach for low snr and non-stationary noise,” in ICASSP 2020- 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE,...

  6. [14]

    Snr- progressive model with harmonic compensation for low-snr speech enhancement,

    Z. Hou, T. Lei, Q. Hu, Z. Cao, M. Tang, and J. Lu, “Snr- progressive model with harmonic compensation for low-snr speech enhancement,” IEEE Signal Processing Letters, 2024

  7. [15]

    Harmonic enhancement using learnable comb filter for light-weight full-band speech en- hancement model,

    X. Le, T. Lei, L. Chen, Y . Guo, C. He, C. Chen, X. Xia, H. Gao, Y . Xiao, P. Ding, S. Song, and J. Lu, “Harmonic enhancement using learnable comb filter for light-weight full-band speech en- hancement model,” in Interspeech 2023, 2023, pp. 3894–3898

  8. [16]

    Speech extraction un- der extremely low snr conditions,

    H. Ruan, L. Liao, K. Chen, and J. Lu, “Speech extraction un- der extremely low snr conditions,” Applied Acoustics, vol. 224, p. 110149, 2024

  9. [17]

    A hybrid dsp/deep learning approach to real-time full-band speech enhancement,

    J.-M. Valin, “A hybrid dsp/deep learning approach to real-time full-band speech enhancement,” in 2018 IEEE 20th international workshop on multimedia signal processing (MMSP) . IEEE, 2018, pp. 1–5

  10. [18]

    Gtcrn: A speech enhancement model requiring ultralow computational resources,

    X. Rong, T. Sun, X. Zhang, Y . Hu, C. Zhu, and J. Lu, “Gtcrn: A speech enhancement model requiring ultralow computational resources,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 971–975

  11. [19]

    DPCRN: Dual-Path Con- volution Recurrent Network for Single Channel Speech Enhance- ment,

    X. Le, H. Chen, K. Chen, and J. Lu, “DPCRN: Dual-Path Con- volution Recurrent Network for Single Channel Speech Enhance- ment,” in Interspeech 2021, 2021, pp. 2811–2815

  12. [20]

    Shufflenet v2: Practi- cal guidelines for efficient cnn architecture design,

    N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “Shufflenet v2: Practi- cal guidelines for efficient cnn architecture design,” in Proceed- ings of the European conference on computer vision (ECCV) , 2018, pp. 116–131

  13. [21]

    Effi- cient sequence learning with group recurrent networks,

    F. Gao, L. Wu, L. Zhao, T. Qin, X. Cheng, and T.-Y . Liu, “Effi- cient sequence learning with group recurrent networks,” in Pro- ceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, V olu...

  14. [22]

    Independent vector analysis: An extension of ica to multivariate components,

    T. Kim, T. Eltoft, and T.-W. Lee, “Independent vector analysis: An extension of ica to multivariate components,” in International conference on independent component analysis and signal sepa- ration. Springer, 2006, pp. 165–172

  15. [23]

    Blind source sep- aration exploiting higher-order frequency dependencies,

    T. Kim, H. T. Attias, S.-Y . Lee, and T.-W. Lee, “Blind source sep- aration exploiting higher-order frequency dependencies,” IEEE transactions on audio, speech, and language processing , vol. 15, no. 1, pp. 70–79, 2006

  16. [24]

    Icassp 2021 deep noise suppres- sion challenge,

    C. K. Reddy, H. Dubey, V . Gopal, R. Cutler, S. Braun, H. Gamper, R. Aichner, and S. Srinivasan, “Icassp 2021 deep noise suppres- sion challenge,” in ICASSP 2021-2021 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2021, pp. 6623–6627

  17. [25]

    Complex ratio mask- ing for monaural speech separation,

    D. S. Williamson, Y . Wang, and D. Wang, “Complex ratio mask- ing for monaural speech separation,” IEEE/ACM transactions on audio, speech, and language processing , vol. 24, no. 3, pp. 483– 492, 2015

  18. [26]

    Stable and fast update rules for independent vector anal- ysis based on auxiliary function technique,

    N. Ono, “Stable and fast update rules for independent vector anal- ysis based on auxiliary function technique,” in 2011 IEEE Work- shop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2011, pp. 189–192

  19. [27]

    An auxiliary-function approach to online independent vector analy- sis for real-time blind source separation,

    T. Taniguchi, N. Ono, A. Kawamura, and S. Sagayama, “An auxiliary-function approach to online independent vector analy- sis for real-time blind source separation,” in 2014 4th Joint Work- shop on Hands-free Speech Communication and Microphone Ar- rays (HSCMA). IEEE, 2014, pp. 107–111

  20. [28]

    Learning cross-lingual sentence representa- tions via a multi-task dual-encoder model,

    M. Chidambaram, Y . Yang, D. Cer, S. Yuan, Y . Sung, B. Strope, and R. Kurzweil, “Learning cross-lingual sentence representa- tions via a multi-task dual-encoder model,” in Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP- 2019). Association for Com...

  21. [29]

    and DNSMOS P.835 [30] are also employed. 5.2. Ablation study We conduct an ablation study on our modified GTCRN to eval- uate the impact of various factors, including the use of speech and noise information from IV A, the type of feature, the type of masking approach, and the ...

  22. [30]

    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,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 46–50

  23. [31]

    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

  24. [32]

    Image method for efficiently sim- ulating small-room acoustics,

    J. B. Allen and D. A. Berkley, “Image method for efficiently sim- ulating small-room acoustics,” The Journal of the Acoustical So- ciety of America, vol. 65, no. 4, pp. 943–950, 1979

  25. [33]

    Description and discussion on dcase 2022 chal- lenge task 2: Unsupervised anomalous sound detection for ma- chine condition monitoring applying domain generalization tech- niques,

    K. Dohi, K. Imoto, N. Harada, D. Niizumi, Y . Koizumi, T. Nishida, H. Purohit, T. Endo, M. Yamamoto, and Y . Kawaguchi, “Description and discussion on dcase 2022 chal- lenge task 2: Unsupervised anomalous sound detection for ma- chine condition monitoring applying domain gener...

  26. [34]

    Adam: A method for stochastic opti- mization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization,” in ICLR 2015, 2015

  27. [35]

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

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,” in 2001 IEEE international conference on acoustics, speech, and signal processing. Proceed...

  28. [36]

    A short- time objective intelligibility measure for time-frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “A short- time objective intelligibility measure for time-frequency weighted noisy speech,” in 2010 IEEE international conference on acous- tics, speech and signal processing. IEEE, 2010, pp. 4214–4217

  29. [37]

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

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

  30. [38]

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

    ——, “Dnsmos p. 835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 886–890

Pith tools

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