Single-Channel Target Speech Extraction Utilizing Distance and Room Clues

Benjamin Yen; Jiang Wang; Kazuhiro Nakadai; Ragib Amin Nihal; Runwu Shi; Zirui Lin

REVIEW 4 major objections 6 minor 27 references

Single-Channel Target Speech Extraction Utilizing Distance and Room Clues

T0 review · 4 major / 6 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 Feeding a speech-extraction model the room's size and echo time improves how well it can pull out a speaker at a given distance.

desk verdict Room-conditioned distance-based TSE is a real, useful extension, but the main generalization claim hinges on an under-specified dataset split that needs clarification before the paper can be fully trusted. read the letter →

arxiv 2505.14433 v1 pith:WBQYWMZY submitted 2025-05-20 eess.AS cs.SD

Runwu Shi , Zirui Lin , Benjamin Yen , Jiang Wang , Ragib Amin Nihal , Kazuhiro Nakadai This is my paper · ORCID

classification eess.AScs.SD

keywords targetspeechextractiondistance-basedsoundseparationroomacousticsreverberationtimesingle-channellearnableembeddingsgeneralizationdirect-to-reverberationratio

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a single-channel target speech extraction system guided only by the target speaker's distance struggles when the room changes, because the same distance implies different acoustics in different enclosures. It proposes feeding the model two extra clues: the six microphone-to-wall distances, which encode room dimensions and microphone placement, and the reverberation time. With these room clues, extraction quality improves on a 1,000-room simulated test set and on real recorded room impulse responses without finetuning. The intended payoff is a distance-based extractor that generalizes across rooms instead of memorizing one room's acoustics.

What carries the argument

The load-bearing component is the Query Embedding Generator, which maps three clues -- query distance $d_q$, a six-element vector of microphone-to-wall distances, and reverberation time $ ext{RT}_{60}$ -- through separate linear layers into embeddings $E_{\mathrm{dis}}$, $E_{\mathrm{dim}}$, and $E_{\mathrm{rt}}$, concatenates them, and produces a $D$-dimensional query embedding $E_Q$. Query blocks then fuse this embedding with the time-frequency representation along temporal and spectral axes using bidirectional LSTMs, so every TF bin is conditioned on both distance and room context. The six microphone-to-wall distances are summed into a single ambiguous embedding, which the authors say generalizes better than a literal coordinate encoding. This mechanism is what lets the same trained model adapt its notion of "near one meter" to rooms of different sizes and decay times.

What would settle it

Take the trained Sim2 model, feed it room clues deliberately mismatched to the actual room (for instance, reverberation time off by 0.2 seconds or microphone position shifted by 1 meter), and measure SDR on the Sim2 test set; if the gain over the distance-only baseline disappears or reverses under realistic measurement error, the claim that room clues improve practical generalization would not survive, whereas a sweep over error magnitudes would reveal the sensitivity boundary.

Watch

Extended reading notes

Core claim

The central claim is that room environmental information is a usable, effective auxiliary clue for distance-based target speech extraction. In the model, a query embedding generator turns the query distance, the six microphone-to-wall distances, and the reverberation time into learnable embeddings that are fused into the time-frequency representation by query blocks, while the same architecture with only the distance clue serves as the baseline. On a simulated dataset spanning 1,000 random rooms, adding room clues raises SDR from 6.71 dB to 7.96 dB and improves the inactive-source loss, and on a real collected room impulse response dataset from an unseen conference room the gain is 3.06 dB to 4.50 dB. The paper takes these numbers as evidence that the distance clue alone is not room-invariant, and that explicit room context restores much of the lost performance.

Load-bearing premise

The method assumes that exact ground-truth values for the six microphone-to-wall distances and the reverberation time are available at inference, and the paper never tests how performance degrades when these clues are wrong, noisy, or estimated.

Editorial extensions

If this is right

If room clues are this useful, distance-based target speech extraction can be deployed in unseen rooms without per-room finetuning, which is a main blocker for real use.
The gain on the real recorded RIR dataset, from 3.06 to 4.50 dB SDR, suggests the improvement transfers from simulated training to measured acoustics.
Better handling of inactive conditions, reflected in lower $L_0$ loss, implies the model also becomes more reliable at saying "no speaker at this distance," which matters for downstream distance estimation.
The comparison with enrolled-voice baselines sets a practical benchmark: distance-only clues are weaker than a voice enrollment, and adding room context closes part of that gap.

Reading between the lines

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

An untested but natural extension is robustness to imperfect room metadata; if the six microphone-to-wall distances or the reverberation time are estimated rather than measured, errors could shrink the 1.25 dB gain, and the paper's assumption of exact clues needs a sensitivity study.
Because the model already conditions on room geometry and decay time, one could try replacing the ground-truth room clues with estimates derived from the mixture itself, turning the system into a fully self-contained extractor.
The same room-embedding mechanism could transfer to multi-channel distance-plus-direction extraction, where room context would help disambiguate directional cues in reverberant enclosures.
The released real-RIR dataset invites a cross-room benchmark that tests whether the gain persists when training and test rooms differ in size by a wider margin than the simulated 4x5x2.5 to 8x10x3 meter range.

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.

Desk Editor's Note

read the letter

The paper does something genuinely useful: it takes the authors' own distance-only single-channel TSE and conditions it on room geometry (mic-wall distances) and RT60, with learnable embeddings. The ablations are consistent—adding room clues helps on both simulated and real RIRs—and the public GridRIR dataset is a concrete contribution that others can build on. The real-RIR generalization test, where a model trained only on simulated rooms is evaluated on a measured room, is the right probe and the 3.06→4.50 dB SDR gain is the most credible evidence in the paper.

The soft spots are real but not disqualifying. The most serious is the split unit for the Sim2 dataset. The paper says only that RIRs were split 0.9/0.02/0.08, not whether the split was by room or by individual RIR sample. With one microphone position per room, room clues are effectively a room identifier. If the same room appears in both training and test, the Sim2 gain (6.71→7.96 dB) could reflect within-room generalization to new speaker positions, not the cross-room generalization claimed in the abstract. The real-RIR experiment is genuinely cross-room, so the central idea probably holds, but the paper needs to state the split and ideally rerun Sim2 with a room-disjoint split. Without that, the primary quantitative support is ambiguous.

Two smaller issues: there is no random-clue control, so part of the gain could come from added capacity rather than the information content of the room embedding (the parameter increase is tiny, 1.25M→1.29M, which mitigates this, but a control would settle it), and there are no error bars despite five test runs. The real-speech finetuning in Table III compares finetuning strategies but omits a distance-only baseline, so it does not isolate the room-clue benefit on real speech.

The assumption that exact room metadata is available at inference is stated clearly and is a practical limitation, not a hidden flaw—the paper is honest about what it conditions on. Robustness to inaccurate RT60 or mic-wall estimates is untested, but that is a follow-up, not a fatal gap.

Who should read this: anyone working on distance-based TSE, spatial clues for speech separation, or single-channel extraction in reverberant rooms. It deserves a serious referee. I would send it to review with two explicit requests: clarify the Sim2 split unit and, if needed, re-run with a room-disjoint split; and add a random-clue or shuffled-clue control. The dataset alone justifies engagement, and the core idea is sound enough that a revision could make it a solid contribution.

Referee Report

4 major / 6 minor

Summary. The paper proposes a single-channel target speech extraction (TSE) system that conditions on both a query distance and room clues: the six microphone-to-wall distances and the reverberation time RT60. The distance and room clues are converted into learnable embeddings by a query embedding generator and fused into a time-frequency domain network with query and basic blocks. The authors evaluate on simulated RIR datasets (Sim1 with one room, Sim2 with 1,000 rooms) and on a newly collected real RIR dataset, plus real recorded speech after finetuning. The main quantitative results are that adding room clues raises SDR on Sim2 from 6.71 dB to 7.96 dB and on the real RIR test set from 3.06 dB to 4.50 dB without finetuning, with only a 0.04M parameter increase. The paper also releases the real RIR dataset publicly.

Significance. If the cross-room generalization claim holds, this is a useful and relatively underexplored extension of distance-based TSE: it uses non-biometric spatial clues, works with a single microphone, and adds only a small number of parameters. The paper has concrete strengths: the real-RIR out-of-distribution test is an appropriate probe for generalization; the released RIR dataset is a community resource; the comparison with enrolled-voice TSE methods situates the approach honestly; and the finetuning experiments on real recorded speech demonstrate practical feasibility. However, the central quantitative support for cross-room generalization is currently ambiguous because the unit of the train/test split in Sim2 is not specified, and the experimental section lacks variance estimates and a random-clue control. The paper's contribution is potentially solid, but the evidence needs to be tightened before the generalization claim is established.

major comments (4)

[Section III-A, Dataset Generation] Please specify the unit of the 0.9:0.02:0.08 split for the Simulated RIR datasets. The text says only 'Simulated RIR datasets were split into training, validation, and testing datasets with a ratio of 0.9:0.02:0.08.' If the split is at the RIR-sample level rather than the room level, then the same room, with its exact RT60 and six microphone-to-wall distances, appears in both training and test sets. In that case the room clue can act as a room identifier, and the Table I Sim2 gain (6.71 dB to 7.96 dB) would measure within-room generalization to new positions and speakers, not the cross-room generalization claimed in the abstract and conclusion. Please state the split unit explicitly; if the current split is sample-based, rerun the Sim2 experiments with a room-disjoint split.
[Section II-B, Query Embedding Generator, and Tables I-II] The evaluation assumes exact access to the six microphone-to-wall distances and RT60 at inference time. Since the claimed practical benefit is generalization across real rooms, where such metadata must be estimated or measured with error, the paper should include a robustness experiment with perturbed room clues. For example, add noise to RT60 and to each mic-wall distance (e.g., ±10% and ±20%) and show how the SDR in Table II degrades. Without this test, it is unclear whether the reported gains survive realistic input uncertainty, which is a load-bearing assumption for the method's practical value.
[Section IV-A, Results on Simulated RIR Dataset] The paper states that each result is obtained by testing five times and expressed as the mean, but Tables I and II report only the means with no standard deviations or significance tests. The reported differences (1.25 dB on Sim2, 1.44 dB on real RIRs) could be within run-to-run variability, especially given the random generation of 1,000 test utterances per run. Please report the standard deviation or per-run values for the key comparisons, or provide a paired significance test.
[Section II-B and Table IV] To attribute the improvement specifically to the room information, please add a control condition in which the room clue is replaced by a random or permuted vector of the same dimension. The parameter increase from Dis to Dis+Dim+Rt is only 0.04M, which makes a pure capacity effect unlikely, but a random-clue control would directly rule out the possibility that the gain comes from the extra embedding pathway rather than from the semantic content of the room clues. This control is cheap and would strengthen the central claim.

minor comments (6)

[Equation (3)] The notation 'y dq −→ Xk xk(t), ks.t.' is malformed; it should read 'k s.t.' and the expression should be cleaned up so that the conditional extraction is clearly stated.
[Figure 2(c)] The label 'HYInput' in the Query Embedding Generator appears to be a typo; it should read 'HY Input' or simply 'Input'.
[Throughout] The activation function is written inconsistently as 'ReLu' and 'ReLU'; please standardize to 'ReLU'.
[Tables I and II] The column header 'No/o ratio' is not defined in the text. Please spell out what 'No/o' abbreviates (presumably non-overlap/overlap) and state why this ratio is reported.
[Section III-A, Simulated RIR dataset] Sim2 is described as 1,000 rooms with 500 speaker positions each, which would give 500,000 RIRs, but the paper reports 490,325. Please add one sentence explaining the shortfall, for example that smaller rooms cannot accommodate all distance bands.
[Table III] The real recorded speech finetuning results are reported only for the proposed model with room clues. Adding a distance-only baseline under the same finetuning protocol would clarify whether the room clues contribute on real speech, or whether the absolute SDR values should be interpreted only as a feasibility demonstration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: room clues are input features, the real-RIR evaluation is held-out, and self-citations are not load-bearing.

full rationale

This is an empirical machine-learning paper, and its central claim is that adding room environmental information (microphone-to-wall distances and reverberation time) to a distance-based target speech extraction system improves extraction quality and cross-room generalization. The room clues are raw input features used by the query embedding generator, not fitted outputs or parameters estimated from test data. The key quantitative evidence includes a genuinely held-out real-RIR evaluation in Table II, where the real conference room is not among the simulated training rooms, so the reported gain from adding room clues (3.06 dB to 4.50 dB without finetuning) has independent empirical content and is not forced by construction. Self-citations to the authors' prior work, [14] and [16], are used only as architectural or baseline references and do not carry the burden of the new experimental claim. The only potential concern is that the Sim2 train/test split unit is unspecified, which could affect whether Sim2 measures cross-room generalization; however, this is an experimental-validity ambiguity rather than a circularity, because the room clues are still inputs rather than fitted labels, and no equation or construction in the paper makes the reported SDR equivalent to the input clues. The paper does not rename a known result, invoke a uniqueness theorem, or smuggle an ansatz through self-citation. Therefore, no circular step is identified.

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

The central claim relies on the physical acoustics model that a source's distance is encoded in the direct-to-reverberant ratio, on the assumption that six microphone-wall distances plus RT60 sufficiently capture room acoustics, on the availability of exact room metadata at inference, and on the validity of the image-method simulations as a training proxy. The only hand-chosen free parameter that materially affects the definition of the task is rspk.

free parameters (3)

Speaker distance range rspk = 0.5 m for training and Sim tests; 0.1 m for real-RIR finetune
Threshold that defines whether a speaker is present at the query distance. Chosen by hand and changed for finetuning, so it directly influences the task difficulty and all reported SDR values.
Active-loss soft threshold tau = 1e-3
Soft upper bound in the SDR loss (Eq. 5); chosen by hand to stabilize training.
Inactive-loss soft threshold tau_inactive = 1e-2
Soft threshold in the zero-output loss (Eq. 6); chosen by hand. Affects the balance between false acceptance and false rejection.

assumptions (4)

domain assumption The observed mixture is a sum of convolutions of anechoic speech with room impulse responses (Eq. 1-2).
Standard linear system model for reverberant audio; the paper does not justify it, it is assumed.
domain assumption Direct-to-reverberant ratio (DRR) decreases as source-microphone distance increases, making distance a usable clue.
Invoked implicitly in the introduction to motivate distance clues; physical acoustics assumption.
domain assumption Room acoustic state can be summarized by six microphone-wall distances and RT60.
The proposed conditioning uses exactly these inputs; other room properties (geometry details, absorption distribution) are ignored.
domain assumption The randomized image method (pyroomacoustics) generates RIRs representative of real rooms.
Simulated training relies on this; real-RIR results partly validate it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Single-Channel Target Speech Extraction Utilizing Distance and Room Clues." pith.science (2026). https://pith.science/paper/WBQYWMZY

@misc{pith2026250514433,
  author       = {Pith},
  title        = {Pith review of: Single-Channel Target Speech Extraction Utilizing Distance and Room Clues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBQYWMZY}},
  note         = {Machine review of arXiv:2505.14433}
}

read the original abstract

This paper aims to achieve single-channel target speech extraction (TSE) in enclosures utilizing distance clues and room information. Recent works have verified the feasibility of distance clues for the TSE task, which can imply the sound source's direct-to-reverberation ratio (DRR) and thus can be utilized for speech separation and TSE systems. However, such distance clue is significantly influenced by the room's acoustic characteristics, such as dimension and reverberation time, making it challenging for TSE systems that rely solely on distance clues to generalize across a variety of different rooms. To solve this, we suggest providing room environmental information (room dimensions and reverberation time) for distance-based TSE for better generalization capabilities. Especially, we propose a distance and environment-based TSE model in the time-frequency (TF) domain with learnable distance and room embedding. Results on both simulated and real collected datasets demonstrate its feasibility. Demonstration materials are available at https://runwushi.github.io/distance-room-demo-page/.

Figures

Figures reproduced from arXiv: 2505.14433 by the authors.

**Figure 2.** Flowchart of the proposed method showing (a) Overall structure, (b) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** RIR recording locations and environment. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 24 canonical work pages

[1]

Target Speaker Extraction by Directly Exploiting Contextual Information in the Time-Frequency Domain,

X. Yang, C. Bao, J. Zhou, and X. Chen, “Target Speaker Extraction by Directly Exploiting Contextual Information in the Time-Frequency Domain,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2024, pp. 10 476– 10 480

work page 2024
[2]

Coarse-to-Fine Target Speaker Ex- traction Based on Contextual Information Exploitation,

X. Yang, C. Bao, and X. Chen, “Coarse-to-Fine Target Speaker Ex- traction Based on Contextual Information Exploitation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 32, pp. 3795–3810, 2024

work page 2024
[3]

Target Speaker Extraction with Ultra-Short Reference Speech by VE-VE Framework,

L. Yang, W. Liu, L. Tan, J. Yang, and H.-G. Moon, “Target Speaker Extraction with Ultra-Short Reference Speech by VE-VE Framework,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5

work page 2023
[4]

Separate in the Speech Chain: Cross-Modal Conditional Audio-Visual Target Speech Extraction,

Z. Mu and X. Yang, “Separate in the Speech Chain: Cross-Modal Conditional Audio-Visual Target Speech Extraction,” 2024

work page 2024
[5]

Speaker Extraction With Co-Speech Gestures Cue,

Z. Pan, X. Qian, and H. Li, “Speaker Extraction With Co-Speech Gestures Cue,” IEEE Signal Processing Letters, vol. 29, pp. 1467–1471, 2022

work page 2022
[6]

Computer- steered microphone arrays for sound transduction in large rooms,

J. L. Flanagan, J. D. Johnston, R. Zahn, and G. W. Elko, “Computer- steered microphone arrays for sound transduction in large rooms,” The Journal of the Acoustical Society of America , vol. 78, no. 5, pp. 1508– 1518, 1985

work page 1985
[7]

Neural Target Speech Extraction: An overview,

K. Zmolikova, M. Delcroix, T. Ochiai, K. Kinoshita, J. ˇCernock´y, and D. Yu, “Neural Target Speech Extraction: An overview,” IEEE Signal Processing Magazine, vol. 40, no. 3, pp. 8–29, 2023

work page 2023
[8]

Neural Spatial Filter: Target Speaker Speech Separation Assisted with Directional Information,

R. Gu, L. Chen, S.-X. Zhang, J. Zheng, Y . Xu, M. Yu, D. Su, Y . Zou, and D. Yu, “Neural Spatial Filter: Target Speaker Speech Separation Assisted with Directional Information,” in Interspeech 2019 . ISCA, 2019, pp. 4290–4294

work page 2019

Show all 27 references

[9]

Multi-Modal Multi-Channel Target Speech Separation,

R. Gu, S.-X. Zhang, Y . Xu, L. Chen, Y . Zou, and D. Yu, “Multi-Modal Multi-Channel Target Speech Separation,” IEEE Journal of Selected Topics in Signal Processing , vol. 14, no. 3, pp. 530–541, 2020

2020
[10]

Adversarial disentanglement of speaker repre- sentation for attribute-driven privacy preservation,

P.-G. No ´e, M. Mohammadamini, D. Matrouf, T. Parcollet, A. Nautsch, and J.-F. Bonastre, “Adversarial disentanglement of speaker repre- sentation for attribute-driven privacy preservation,” arXiv preprint arXiv:2012.04454, 2020

2012 arXiv
[11]

Sound Source Distance Estimation in Diverse and Dynamic Acoustic Conditions,

S. S. Kushwaha, I. R. Roman, M. Fuentes, and J. P. Bello, “Sound Source Distance Estimation in Diverse and Dynamic Acoustic Conditions,” in 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2023, pp. 1–5

2023
[12]

Speaker Distance Estimation in Enclosures From Single-Channel Au- dio,

M. Neri, A. Politis, D. A. Krause, M. Carli, and T. Virtanen, “Speaker Distance Estimation in Enclosures From Single-Channel Au- dio,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 32, pp. 2242–2254, 2024

2024
[13]

Distance-Based Sound Separation,

K. Patterson, K. Wilson, S. Wisdom, and J. R. Hershey, “Distance-Based Sound Separation,” in Interspeech 2022. ISCA, 2022, pp. 901–905

2022
[14]

Rezero: Region-customizable sound extraction,

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

2024
[15]

Focus on the sound around you: Monaural target speaker extraction via distance and speaker information,

J. Lin, P. Wang, H. Dinkel, J. Chen, Z. Wu, Z. Yan, Y . Wang, J. Zhang, and Y . Wang, “Focus on the sound around you: Monaural target speaker extraction via distance and speaker information,” arXiv preprint arXiv:2306.16241, 2023

2023 arXiv
[16]

Distance based single-channel target speech extraction,

R. Shi, B. Yen, and K. Nakadai, “Distance based single-channel target speech extraction,” arXiv preprint arXiv:2412.20144 , 2024

2024 arXiv
[17]

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. Watanabe, “TF-GridNet: Integrating Full- and Sub-Band Modeling for Speech Separation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 3221–3236, 2023

2023
[18]

Learning neural acoustic fields,

A. Luo, Y . Du, M. Tarr, J. Tenenbaum, A. Torralba, and C. Gan, “Learning neural acoustic fields,” Advances in Neural Information Processing Systems, vol. 35, pp. 3165–3177, 2022

2022
[19]

Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech separation,

Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech separation,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2020, pp. 46–50

2020
[20]

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 International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 626–630

2019
[21]

Soundbeam: Target sound extraction conditioned on sound- class labels and enrollment clues for increased performance and continu- ous learning,

M. Delcroix, J. B. V ´azquez, T. Ochiai, K. Kinoshita, Y . Ohishi, and S. Araki, “Soundbeam: Target sound extraction conditioned on sound- class labels and enrollment clues for increased performance and continu- ous learning,” IEEE/ACM Transactions on Audio, Speech, and Langua...

2022
[22]

What’s all the Fuss about Free Universal Sound Separation Data?

S. Wisdom, H. Erdogan, D. P. W. Ellis, R. Serizel, N. Turpault, E. Fonseca, J. Salamon, P. Seetharaman, and J. R. Hershey, “What’s all the Fuss about Free Universal Sound Separation Data?” in ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Proc...

2021
[23]

Pyroomacoustics: A Python package for audio room simulations and array processing algorithms,

R. Scheibler, E. Bezzam, and I. Dokmani ´c, “Pyroomacoustics: A Python package for audio room simulations and array processing algorithms,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 351–355

2018
[24]

Advancements in impulse response measurements by sine sweeps,

A. Farina, “Advancements in impulse response measurements by sine sweeps,” in Audio engineering society convention 122 . Audio Engi- neering Society, 2007

2007
[25]

Libri-light: A benchmark for asr with limited or no supervision,

J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu, P.-E. Mazar ´e, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegen et al., “Libri-light: A benchmark for asr with limited or no supervision,” in ICASSP 2020- 2020 IEEE International Conference on Acoustics, Speech and Sig...

2020
[26]

Adam: A method for stochastic optimization,

D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

2014 arXiv
[27]

Spex+: A complete time domain speaker extraction network,

M. Ge, C. Xu, L. Wang, E. S. Chng, J. Dang, and H. Li, “Spex+: A complete time domain speaker extraction network,” arXiv preprint arXiv:2005.04686, 2020

2005 arXiv

Pith tools

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

[1] [1]

Target Speaker Extraction by Directly Exploiting Contextual Information in the Time-Frequency Domain,

X. Yang, C. Bao, J. Zhou, and X. Chen, “Target Speaker Extraction by Directly Exploiting Contextual Information in the Time-Frequency Domain,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2024, pp. 10 476– 10 480

work page 2024

[2] [2]

Coarse-to-Fine Target Speaker Ex- traction Based on Contextual Information Exploitation,

X. Yang, C. Bao, and X. Chen, “Coarse-to-Fine Target Speaker Ex- traction Based on Contextual Information Exploitation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 32, pp. 3795–3810, 2024

work page 2024

[3] [3]

Target Speaker Extraction with Ultra-Short Reference Speech by VE-VE Framework,

L. Yang, W. Liu, L. Tan, J. Yang, and H.-G. Moon, “Target Speaker Extraction with Ultra-Short Reference Speech by VE-VE Framework,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5

work page 2023

[4] [4]

Separate in the Speech Chain: Cross-Modal Conditional Audio-Visual Target Speech Extraction,

Z. Mu and X. Yang, “Separate in the Speech Chain: Cross-Modal Conditional Audio-Visual Target Speech Extraction,” 2024

work page 2024

[5] [5]

Speaker Extraction With Co-Speech Gestures Cue,

Z. Pan, X. Qian, and H. Li, “Speaker Extraction With Co-Speech Gestures Cue,” IEEE Signal Processing Letters, vol. 29, pp. 1467–1471, 2022

work page 2022

[6] [6]

Computer- steered microphone arrays for sound transduction in large rooms,

J. L. Flanagan, J. D. Johnston, R. Zahn, and G. W. Elko, “Computer- steered microphone arrays for sound transduction in large rooms,” The Journal of the Acoustical Society of America , vol. 78, no. 5, pp. 1508– 1518, 1985

work page 1985

[7] [7]

Neural Target Speech Extraction: An overview,

K. Zmolikova, M. Delcroix, T. Ochiai, K. Kinoshita, J. ˇCernock´y, and D. Yu, “Neural Target Speech Extraction: An overview,” IEEE Signal Processing Magazine, vol. 40, no. 3, pp. 8–29, 2023

work page 2023

[8] [8]

Neural Spatial Filter: Target Speaker Speech Separation Assisted with Directional Information,

R. Gu, L. Chen, S.-X. Zhang, J. Zheng, Y . Xu, M. Yu, D. Su, Y . Zou, and D. Yu, “Neural Spatial Filter: Target Speaker Speech Separation Assisted with Directional Information,” in Interspeech 2019 . ISCA, 2019, pp. 4290–4294

work page 2019