Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

End-to-End DOA-Guided Speech Extraction in Noisy Multi-Talker Scenarios

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Cyclic DOA embeddings let a compact neural beam isolate one speaker among six.

desk verdict Reasonable DOA-conditioned extraction front end, but the headline numbers come from the Stage 1 model with the beamwidth module disabled, so the central claim about beamwidth-guided extraction is not actually measured. read the letter →

arxiv 2507.20926 v1 pith:63IJNKUB submitted 2025-07-28 eess.AS

Kangqi Jing , Wenbin Zhang , Yu Gao This is my paper
classification eess.AS
keywords targetspeakerextractiondirectionalspeechdirectionofarrivalembeddingbeamwidthcontrolmulti-talkerseparationautomaticrecognitioncyclicpositionalencodingneuralbeamforming
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 an end-to-end neural model can use a direction-of-arrival (DOA) value and a beamwidth as soft spatial cues to extract the speech of one chosen speaker from a reverberant mixture containing up to six simultaneous talkers plus noise. The authors argue that cyclic positional DOA embeddings and a learned beamwidth mask let the network form an adaptive beam: it boosts signals inside the chosen angular region and suppresses directions outside it. If the reported results transfer, such a compact 1.4M-parameter model would make directional speech extraction practical for smart devices and conference systems, and would substantially improve downstream automatic speech recognition in multi-talker conditions.

What carries the argument

The central object is the cyclic positional DOA embedding $\mathrm{PE}_{\mathrm{cyc\text{-}pos}}(\phi)$ from Eq. 2, which maps an angle to a vector by composing $\sin(\phi)$ and $\cos(\phi)$ with frequency-scaled position codes; this avoids the one-hot discontinuity between 359 and 0 degrees and shrinks the embedding from 360 dimensions to 40. The second mechanism is the BW-ConvModule, a learned mask generated from a one-hot beamwidth vector through a linear layer and a $1\times1$ convolution; it multiplies the feature map to suppress directions outside the requested beam. The DOA embedding is broadcast in time, refined by a clue encoder (linear + layer norm + PReLU), and applied by element-wise multiplication to the main feature stream. The stream itself alternates Crossband layers, which model frequency correlations, and Narrowband layers, which apply multihead self-attention per frequency to separate sources by direction.

What would settle it

Re-run the trained model with the input DOA deliberately offset by 5, 10, or 20 degrees from the true speaker angle, or feed it the output of a blind DOA estimator on the same mixtures; if SI-SDR improvement falls by more than a few dB or the gain pattern's peak shifts off the true speaker, the claimed directional selectivity and beamwidth tolerance are not robust.

Watch

Extended reading notes

Core claim

This paper demonstrates that an end-to-end network can extract one speaker's voice from a reverberant, noisy mixture of up to six simultaneous talkers by conditioning the network on a target direction and a beamwidth. Its best configuration uses a 40-dimensional cyclic positional DOA embedding with scaling factor 20 and reaches 18.29 dB SI-SDR improvement, 13.99 dB SDR improvement, and PESQ 1.40 on the simulated test set, using 1.40M parameters. The same model lowers WER for downstream ASR from 82.10% (noisy) and 38.06% (JNF baseline) to 10.52% on two-speaker mixtures, and to 17.31% on three-speaker mixtures. Gain-pattern analysis shows the model's response peaks inside the requested beam and stays below -20 dB outside it, with 5-10 degrees of spillover that grows with beamwidth.

Load-bearing premise

The demonstration assumes the target DOA handed to the model is exactly right and that real rooms match the simulated set (6-9 m rooms, RT60 0.3-0.5 s, a 3-microphone 30 mm circular array), since the paper never tests perturbed DOAs or real recordings.

Editorial extensions

If this is right

  • A 1.4M-parameter extractor can handle six overlapping speakers in reverberation, a regime where a conventional MVDR beamformer fails outright.
  • A 40-dimensional angle embedding can outperform a 360-dimensional one-hot while using 0.49M fewer parameters, so the choice of angular representation matters more than raw capacity.
  • The beamwidth input acts as a zoom control: wider beams preserve all voices inside them, and a beam with no active speaker produces near-silence rather than noise.
  • Using the extracted voice as an ASR front-end reduces word error rate from 38.06% to 10.52% on two-talker mixtures and from 57.04% to 17.31% for three talkers.
  • Stage-2 training with 10% empty beams is sufficient to teach the model a silence-like output, avoiding the instability of SI-SDR on silent targets.

Reading between the lines

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

  • Testable extension: feed the model a DOA from an actual estimator or add angular jitter during training; the paper's robustness claim is untested because it always used the true direction.
  • Extension: map requested beamwidth to measured gain-pattern width; spillover of 5-10 degrees suggests the effective beam is wider than the setting.
  • Extension: repeat with real room impulse responses or a different microphone array; all evidence comes from one simulated room configuration.
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

5 major / 5 minor

Summary. The paper proposes DSENet, an end-to-end multichannel target speaker extraction model that conditions on a target DOA and a beamwidth through cyclic positional (cyc-pos) embeddings and a BW-ConvModule mask, built on interleaved Crossband/Narrowband blocks adapted from SpatialNet. Training proceeds in two stages: first target extraction with the beamwidth module disabled, then beam adaptation with widths of 15, 30, and 45 degrees and a mixture of active and inactive beams. On a simulated six-speaker reverberant LibriSpeech/DEMAND test set, the best Stage 1 configuration achieves 18.29 dB SI-SDRi, 13.99 dB SDRi, and 1.40 PESQ, outperforming MVDR and JNF. On separate two- and three-speaker ASR tests, the paper reports 10.52% and 17.31% WER, respectively. Gain-pattern plots and per-DOA extraction results are used to support the beamwidth behavior.

Significance. If the reported numbers are reproducible, the contribution is a compact (1.40M parameter) neural beamformer that combines DOA and explicit beamwidth control, with a qualitatively convincing gain pattern showing a narrow passband and strong rejection outside the beam. The two-stage training protocol, including the use of a low-level reference tone for inactive beams, is an interesting practical solution to a real training instability. The availability of code and audio examples is a strength. However, the paper's quantitative evaluation currently bypasses the very beamwidth module that constitutes the main architectural contribution, and several evaluation choices (test-set selection of alpha, true-DOA-only testing, sparse baselines) limit the strength of the claims.

major comments (5)
  1. [Section 4.1, Table 1] The headline objective metrics (SI-SDRi 18.29 dB, SDRi 13.99 dB, PESQ 1.40) are obtained with the Stage 1 model, in which the BW-ConvModule is explicitly disabled. The central claim of the paper concerns the full DOA-plus-beamwidth system, but no objective metric is reported for the Stage 2 model with the BW-ConvModule enabled. Table 2 does not state which stage is used for the ASR results, making them ambiguous as well. Please report SI-SDRi/SDRi/PESQ/WER for the full Stage 2 model across the three beamwidths, and include an on/off ablation of the beamwidth module so the reader can determine whether the module improves, preserves, or degrades extraction quality.
  2. [Section 4.1, Table 1] The cyclic positional scaling factor alpha is selected on the same test set used to report the headline result: Table 1 reports alpha = 10, 20, and 40 and the best value (alpha = 20) is then used for the comparisons. This constitutes test-set model selection and biases the reported improvement. Please fix alpha on a held-out validation set, or describe the selection procedure explicitly and, ideally, report error bars across mixture draws.
  3. [Sections 3.1 and 4.1] The evaluation always feeds the true speaker DOA to the model. The introduction identifies sensitivity to inaccurate DOA as the key limitation of prior work, and the beamwidth is claimed to provide tolerance to DOA estimation errors, but no experiment tests this. Section 4.1 says 'we set the input DOA such that only one speaker falls within the beam width,' and Section 4.3 similarly sweeps the input DOA using the true speaker directions. Please add an experiment with perturbed DOA inputs (e.g., errors of +/-5, +/-10, and +/-15 degrees) and report extraction metrics as a function of the error.
  4. [Sections 2.1, 4.1, and 4.4] The claimed behavior of retaining multiple speakers within a beam, and of producing silence when the beam is inactive, is not quantitatively evaluated. Section 4.1 restricts evaluation to one speaker inside the beam, and Section 4.4 only says the input DOA was adjusted to the range of active sound sources without specifying whether the beam contains one or all speakers. The paper should evaluate mixtures with two speakers inside the same beam and with inactive beams, using a suitable metric for the silence/tone output.
  5. [Section 4.1 and Table 2] The comparison set is thin: only MVDR and JNF are included, with no modern DSE baselines such as ReZero [13] or all-neural low-latency directional speech extraction [14], and no variance or statistical significance measures are reported. Given the claim that the model 'significantly outperforms' JNF, the absence of error bars or a broader baseline suite makes the magnitude of the improvement hard to assess.
minor comments (5)
  1. [Section 4.1] There are typos: 'trained int Stage 1' should be 'trained in Stage 1', and 'reference channal' in Section 4.2 should be 'reference channel'.
  2. [Sections 2.2.1 and 2.3] The symbol alpha is used both for the cyc-pos scaling factor in Eq. (2) and for the optimal scaling factor in Eqs. (5)-(6); please rename one of them to avoid confusion.
  3. [Section 4.4 and Table 2] The ASR setup is underspecified: the ASR model, training data, and decoding details are not given, and the first column of Table 2 has no header. The label 'Mixed Noisy' should be clarified (e.g., unprocessed mixture), and the paper should state whether the ASR results use the Stage 1 or Stage 2 model.
  4. [Section 4.2, Figure 2] The gain-pattern figures would be clearer with an explicit caption describing the dB scale, the microphone array geometry, the room configuration, and the number of sources used to generate each pattern; the current radial scale is hard to read.
  5. [Section 2.1] The beamwidth notation is inconsistent: the beam is defined as [theta_target - theta_width, theta_target + theta_width], but the input is called theta_beam and the module takes a 'width input'; please standardize the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the model is evaluated on independent simulated test mixtures, and the Stage-1-only headline metrics are a support gap rather than a circular step.

full rationale

I followed the paper's derivation chain from the problem formulation (Section 2.1), through the architecture and loss (Sections 2.2-2.3), to the experimental protocol (Sections 3-4). The model is not derived from its target; it is trained with an SI-SDR-plus-magnitude loss on simulated mixtures and then evaluated on held-out simulated mixtures, so the reported SI-SDRi, SDRi, PESQ, and WER values are genuine empirical outputs rather than restatements of the inputs. The only passage that deserves scrutiny is Section 4.1: "For evaluation fairness and simplicity, we use the model trained in Stage 1, where the BW-ConvModule is disabled, and set the input DOA such that only one speaker falls within the beam width." This means Table 1's numbers do not measure the full BW-ConvModule system, and Figure 2's gain patterns are the only evidence for the beamwidth module's behavior. That is a missing-support or reporting issue, not circularity: the Stage-1 model is still evaluated against external test data, and the beamwidth module is not a fitted parameter being renamed as a prediction. Likewise, the selection of alpha=20 in Table 1 as the best-scoring cyc-pos configuration is test-set hyperparameter selection, which is a methodological concern, not a derivation that assumes its conclusion. I found no load-bearing self-citation, no author-imported uniqueness theorem, no ansatz smuggled in through prior work, and no renamed known result. The task definition uses the oracle DOA as the conditioning input, so strong performance given that DOA is an empirical finding rather than a tautology.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central contribution is a trained neural network; there are no new physical entities. The main hand-set quantities are the cyc-pos scale alpha, chosen on validation and test performance, and the loss weights. Assumptions are the standard additive mixture model and the validity of simulated room acoustics. The model reuses Crossband and Narrowband blocks from prior work, which is a background assumption rather than a new entity.

free parameters (2)
  • cyc-pos scaling factor alpha = 20 (best of {10, 20, 40})
    Chosen by comparing SI-SDRi for alpha=10, 20, 40 in Table 1; alpha=20 is selected and alpha=40 degrades performance, so the reported result depends on this hand-picked hyperparameter.
  • loss weighting lambda = 0.5 (stage 1), 0.05 (stage 2)
    Set by hand; affects optimization and the final model's behavior, including how the model treats inactive beams.
assumptions (5)
  • domain assumption Signal model in Eq. (1): mixture is sum of convolutions of clean sources with room impulse responses plus independent noise.
    All experiments rely on this linear additive mixture model; real recordings may include nonlinearities and diffuse noise.
  • domain assumption Simulated room acoustics with pyroomacoustics are representative of real conditions.
    Section 3.1 specifies RT60 0.3-0.5s, room 6-9m, 3-microphone circular array; generalization to real rooms is untested.
  • domain assumption DOA input is accurate or oracle during evaluation.
    Section 4.1 sets input DOA to the true speaker DOA; robustness to DOA estimation errors is not measured.
  • domain assumption The reused Crossband and Narrowband layers from SpatialNet [19] are effective for multichannel speech separation as claimed in that prior work.
    The paper builds its backbone directly on [19] without re-validating the component independently, relying on the cited result.
  • ad hoc to paper A 20 Hz reference tone at -60 dB RMS represents silence for inactive beams during Stage 2 training.
    Section 3.2 uses this tone to avoid SI-SDR instability; it changes the learning target for inactive cases and may affect behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-End DOA-Guided Speech Extraction in Noisy Multi-Talker Scenarios." pith.science (2026). https://pith.science/paper/63IJNKUB

@misc{pith2026250720926,
  author       = {Pith},
  title        = {Pith review of: End-to-End DOA-Guided Speech Extraction in Noisy Multi-Talker Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63IJNKUB}},
  note         = {Machine review of arXiv:2507.20926}
}
read the original abstract

Target Speaker Extraction (TSE) plays a critical role in enhancing speech signals in noisy and multi-speaker environments. This paper presents an end-to-end TSE model that incorporates Direction of Arrival (DOA) and beamwidth embeddings to extract speech from a specified spatial region centered around the DOA. Our approach efficiently captures spatial and temporal features, enabling robust performance in highly complex scenarios with multiple simultaneous speakers. Experimental results demonstrate that the proposed model not only significantly enhances the target speech within the defined beamwidth but also effectively suppresses interference from other directions, producing a clear and isolated target voice. Furthermore, the model achieves remarkable improvements in downstream Automatic Speech Recognition (ASR) tasks, making it particularly suitable for real-world applications.

Figures

Figures reproduced from arXiv: 2507.20926 by the authors.

Figure 1
Figure 1. (a) The system overview of the proposed DSE model. The detailed structure of (b) Clue encoder, (c) BW-ConvModule, (d) Crossband Layer and (e) Narrowband Layer. retain and output all voices within this range. If no speech is present in the range, the model outputs a signal that is either close to silence or silent. 2.2. Model Architecture The overview of the proposed architecture is depicted in Fig￾ure 1. First, a mu… view at source ↗
Figure 3
Figure 3. SI-SDRi for six speakers under various input DOAs with a fixed beamwidth of 30°, The DOAs of the six speakers (Spk1-Spk6) are: 0°, 50°, 162°, 187°, 214°, and 313°. This work primarily evaluates the model’s ability to extract clear speech signals corresponding to different speakers under varying input DOAs. To this end, we calculate the SI-SDRi for each speaker based on the model’s output, with the beamwidth input fi… view at source ↗
Figure 2
Figure 2. Gain-pattern analysis for different input beamwidths: (a) 15°, (b) 30°, and (c) 45°. Two random DOA inputs are tested for each beamwidth, with the input DOA indicated by a red dot. speech extraction range and offers a tolerance for DOA estima￾tion errors, making it more aligned with practical requirements. 4.3. Effective Speaker Extraction Capability 0 30 60 90 120 150 180 210 240 270 300 330 input doa (°) 16 18 20 … view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. WeSep: A Modular and Cue-Composable Framework for Target Speaker Extraction

    eess.AS 2026-07 conditional novelty 4.0 of 10

    WeSep decouples cue frontends from separator backbones via standardized interfaces and shows stable multi-cue and missing-cue TSE training across enrollment, spatial, visual, and textual modalities.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [13]

    Robust Target Speaker Direction of Arrival Estimation

    Z. Li, S. He, and X. Zhang, “Robust target speaker direction of arrival estimation,” arXiv preprint arXiv:2412.18913, 2024

    work page Pith review arXiv 2024

  2. [14]

    Dsenet: Directional sig- nal extraction network for hearing improvement on edge devices,

    A. Kovalyov, K. Patel, and I. Panahi, “Dsenet: Directional sig- nal extraction network for hearing improvement on edge devices,” IEEE Access, vol. 11, pp. 4350–4358, 2023

    work page 2023

  3. [1]

    cocktail party effect

    Introduction In complex auditory environments with multiple sound sources, humans are able to selectively focus on a specific sound, a phe- nomenon commonly known as the “cocktail party effect”. This ability allows us to attend to a target sound using various cues, such as its time-frequency pattern or the direction of arrival (DOA) [1]. Target Sound Extr...

    work page

  4. [2]

    Problem formulation Let X ∈ RM ×L be the multichannel mixture recorded by an M-microphone array, whereL denotes the number of time sam- ples

    Method 2.1. Problem formulation Let X ∈ RM ×L be the multichannel mixture recorded by an M-microphone array, whereL denotes the number of time sam- ples. The signal at the m-th microphone can be expressed as: Xm(t) = NX i=1 si(t) ∗ Hm i (t) +Nm(t), (1) where si(t) ∈ RL represents the clean speech signal of the i-th speaker, Hm i (t) denotes the room impul...

    arXiv 2025

  5. [3]

    Dataset For this study, the speech data is sourced from the LibriSpeech corpus [23], while the background noise is taken from the DE- MAND dataset [24]

    Experiment 3.1. Dataset For this study, the speech data is sourced from the LibriSpeech corpus [23], while the background noise is taken from the DE- MAND dataset [24]. The room acoustics, including reverber- ation and microphone array configuration, are simulated using the pyroomacoustics1 [25] package. The room’s width and depth are randomly sampled fro...

    work page

  6. [4]

    Results and Analysis 4.1. Ablation Study and Comparison results To evaluate our proposed method’s effectiveness, we compute three metrics: SDR improvement (SDRi) [27], SI-SDR im- provement (SI-SDRi), and perceptual evaluation of speech qual- ity (PESQ) [28]. We also analyzed the parameter count (Para.) to assess computational efficiency. For evaluation fa...

    work page

  7. [5]

    Conclusion In this work, we propose a novel end-to-end target speaker ex- traction model that leverages DOA and beamwidth as soft con- straints to form an adaptive neural beam, dynamically focus- ing on the target speech even in highly complex multi-speaker environments. By integrating DOA and beamwidth embed- dings, our approach efficiently captures spat...

    work page

  8. [6]

    The cocktail party problem,

    S. Haykin and Z. Chen, “The cocktail party problem,” Neural computation, vol. 17, no. 9, pp. 1875–1902, 2005

    1902

Show all 36 references
  1. [7]

    Music source separation with band-split rnn,

    Y . Luo and J. Yu, “Music source separation with band-split rnn,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 31, pp. 1893–1901, 2023

    1901

  2. [8]

    Distance- based sound separation,

    K. Patterson, K. Wilson, S. Wisdom, and J. R. Hershey, “Distance- based sound separation,” inInterspeech 2022, 2022, pp. 901–905

    2022

  3. [9]

    Neural directed speech enhancement with dual microphone array in high noise scenario,

    W. Wen, Q. Zhou, Y . Xi, H. Li, Z. Gong, and K. Yu, “Neural directed speech enhancement with dual microphone array in high noise scenario,” arXiv preprint arXiv:2412.18141, 2024

    2024 arXiv

  4. [10]

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

    S. Wang, X. Kong, X. Peng, H. Movassagh, V . Prakash, and Y . Lu, “Dasformer: Deep alternating spectrogram transformer for multi/single-channel speech separation,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023...

    2023

  5. [11]

    Tf-gridnet: Integrating full-and sub-band modeling for speech separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watan- abe, “Tf-gridnet: Integrating full-and sub-band modeling for speech separation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    2023

  6. [12]

    Deft-an: Dense frequency-time atten- tive network for multichannel speech enhancement,

    D. Lee and J.-W. Choi, “Deft-an: Dense frequency-time atten- tive network for multichannel speech enhancement,”IEEE Signal Processing Letters, vol. 30, pp. 155–159, 2023

    2023

  7. [15]

    Insights into deep non-linear filters for improved multi-channel speech enhancement,

    K. Tesch and T. Gerkmann, “Insights into deep non-linear filters for improved multi-channel speech enhancement,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 563–575, 2022

    2022

  8. [16]

    Neural spatio-temporal beamformer for target speech separa- tion,

    Y . Xu, M. Yu, S.-X. Zhang, L. Chen, C. Weng, J. Liu, and D. Yu, “Neural spatio-temporal beamformer for target speech separa- tion,” in Proc. Interspeech 2020, 2020, pp. 56–60

    2020

  9. [17]

    Multi-channel speech separation us- ing spatially selective deep non-linear filters,

    K. Tesch and T. Gerkmann, “Multi-channel speech separation us- ing spatially selective deep non-linear filters,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 32, pp. 542–553, 2023

    2023

  10. [18]

    Rezero: Region-customizable sound extrac- tion,

    R. Gu and Y . Luo, “Rezero: Region-customizable sound extrac- tion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  11. [19]

    All neural low-latency directional speech extraction,

    A. Pandey, S. Lee, J. Azcarreta, D. Wong, and B. Xu, “All neural low-latency directional speech extraction,” in Proc. Interspeech 2024, 2024, pp. 4328–4332

    2024

  12. [20]

    Direction of arrival correction through speech quality feedback,

    C. Rascon, “Direction of arrival correction through speech quality feedback,” Digital Signal Processing, vol. 158, p. 104960, 2025

    2025

  13. [21]

    Spatio-temporal outdoor lighting aggregation on image se- quences using transformer networks,

    H. Lee, C. Homeyer, R. Herzog, J. Rexilius, and C. Rother, “Spatio-temporal outdoor lighting aggregation on image se- quences using transformer networks,” International Journal of Computer Vision, vol. 131, no. 4, pp. 1060–1072, 2023

    2023

  14. [22]

    Layer normalization,

    J. Lei Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” ArXiv e-prints, pp. arXiv–1607, 2016

    2016

  15. [23]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,” in Proceedings of the IEEE international conference on computer vision, 2015, pp. 1026–1034

    2015

  16. [24]

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

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

    2024

  17. [25]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural In- formation Processing Systems, 2017

    2017

  18. [26]

    Group normalization,

    Y . Wu and K. He, “Group normalization,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 3– 19

    2018

  19. [27]

    Sdr – half-baked or well done?

    J. L. Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “Sdr – half-baked or well done?” in ICASSP 2019 - 2019 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 626–630

    2019

  20. [28]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  21. [29]

    The diverse environments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings,

    J. Thiemann, N. Ito, and E. Vincent, “The diverse environments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings,” in Proceedings of Meetings on Acoustics, vol. 19, no. 1. AIP Publishing, 2013

    2013

  22. [30]

    Pyroomacoustics: A python package for audio room simulation and array processing algorithms,

    R. Scheibler, E. Bezzam, and I. Dokmani ´c, “Pyroomacoustics: A python package for audio room simulation and array processing algorithms,” in 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 351– 355

    2018

  23. [31]

    Adam: A method for stochastic optimiza- tion,

    D. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” Computer Science, 2014

    2014

  24. [32]

    Performance measure- ment in blind audio source separation,

    E. Vincent, R. Gribonval, and C. F ´evotte, “Performance measure- ment in blind audio source separation,” IEEE transactions on au- dio, speech, and language processing , vol. 14, no. 4, pp. 1462– 1469, 2006

    2006

  25. [33]

    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...

    2001

  26. [34]

    A signal subspace tracking algorithm for microphone array processing of speech,

    S. Affes and Y . Grenier, “A signal subspace tracking algorithm for microphone array processing of speech,” IEEE Transactions on Speech and Audio Processing, vol. 5, no. 5, pp. 425–437, 1997

    1997

  27. [35]

    Spatially selective deep non-linear filters for speaker extraction,

    K. Tesch and T. Gerkmann, “Spatially selective deep non-linear filters for speaker extraction,” in ICASSP 2023-2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  28. [36]

    On the importance of acoustic reflections in beamforming,

    O. Shmaryahu and S. Gannot, “On the importance of acoustic reflections in beamforming,” in 2022 International Workshop on Acoustic Signal Enhancement (IWAENC). IEEE, 2022, pp. 1–5

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.