Distance Based Single-Channel Target Speech Extraction

Benjamin Yen; Kazuhiro Nakadai; Runwu Shi

REVIEW 3 major objections 5 minor 1 cited by

Distance Based Single-Channel Target Speech Extraction

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Using only a requested distance as the cue, a single microphone can extract a target speaker's voice from a mixed recording, with no voiceprint or other physiological information.

desk verdict A plausible first demonstration that single-channel TSE can be driven by a scalar distance query, though the evaluation could be stronger. read the letter →

arxiv 2412.20144 v1 pith:UP5UPMBT submitted 2024-12-28 eess.AS cs.SD

Runwu Shi , Benjamin Yen , Kazuhiro Nakadai This is my paper · ORCID

classification eess.AScs.SD

keywords targetspeechextractionsingle-channelsourceseparationdistance-basedsounddirect-to-reverberationratiodistanceembeddingspeakerestimationreverberationtime-frequencymasking

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 argues that a single microphone can extract the speech of a target talker from an overlapping mixture using only a query distance as the cue, with no voiceprint, face, or other physiological information about the speaker. If the argument holds, speech separation becomes possible from a simple physical quantity that is less privacy-sensitive than biometric speaker clues. The authors propose a model that converts the query distance into a learned embedding, fuses it with time-frequency representations in a parameter-efficient Distance Query block, and outputs a mask for speech within a specified distance range, or zero when no talker is near that distance. Experiments show the method works in a fixed simulated room (12.47 dB signal-to-distortion ratio, SDR), degrades when rooms and microphone positions vary (6.83 dB) and on real impulse responses (4.79 dB), and can also estimate speaker distances by scanning the query distance.

What carries the argument

The load-bearing object is the Distance Query (DQ) block, which consists of two learnable distance embedding generators plus BLSTM-based intra-subband and intra-frame fusion modules. It converts the scalar query distance $d_q$ into two embeddings $D_s$ and $D_f$ that are concatenated with the time-frequency embeddings of the mixture and processed with a recurrent layer shared across all subbands or frames, so the network learns which TF bins belong to a talker at the queried distance. The Temporal and Spectral (TS) blocks refine the representation, and the active/inactive SDR losses jointly train the model to produce correct scale when a target is present and silence when it is not.

What would settle it

Train the model on D1 and evaluate it on anechoic or heavily damped recordings where the direct-to-reverberation ratio no longer varies systematically with distance; if extraction accuracy does not drop toward the no-separation baseline, the model is not actually using the distance cue. A complementary check is to move a single speaker along a circle of constant distance from the microphone and confirm that extraction stays stable despite completely different directions and room impulse responses.

Watch

Extended reading notes

Core claim

The paper's central claim is that the distance between a speaker and a single microphone is itself a sufficient clue for target speech extraction, because in an enclosed room the direct-to-reverberation ratio changes systematically with distance. The paper argues that distance is an ambiguous cue—the same distance can correspond to many room impulse responses from different directions—and that this ambiguity is exactly what makes a single-channel solution feasible. Given a query distance $d_q$, the model is trained to output the speech of every speaker whose distance $d_k$ satisfies $|d_k - d_q| \le r_{\mathrm{spk}}$, and to output zero when no speaker satisfies that condition. The same trained model can be used for distance estimation by sweeping the query distance and looking for peaks in the inactive-SDR indicator.

Load-bearing premise

The load-bearing premise is that a speaker's distance from the microphone leaves a consistent acoustic fingerprint in a single-channel mixture — mainly the direct-to-reverberation ratio — and that the distance embedding learned from simulated rooms transfers to unseen rooms, microphone positions, and real impulse responses.

Editorial extensions

If this is right

A distance query can replace an enrollment utterance or a face as the extraction clue, removing the need for biometric speaker data.
Scanning the query distance turns the same model into a speaker-distance estimator, with mean absolute errors of 0.26 m in a fixed room, 0.32 m with random microphone positions, 0.41 m in random rooms, and 1.09 m on real impulse responses.
The trained model's zero-output behavior when no talker sits at the query distance makes it a distance-conditioned presence detector as well as a separator.
At 1.25M parameters, the model is small enough that distance-conditioned extraction could run on devices with tight memory budgets.

Reading between the lines

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

A natural downstream use is an interpretable hearing-aid or recording control: the user specifies a distance ('the person about two meters away') instead of enrolling a voice.
The same iSDR-scanning procedure could be turned into a single-channel range finder that reports how many talkers are active at each distance, with no microphone array.
Because distance is direction-ambiguous, adding a coarse direction cue from a second microphone might shrink the real-room generalization gap reported for D4; this is a plausible extension the paper does not test.
The model's learned distance embedding may be capturing room-specific statistics rather than universal DRR, so a cross-room stress test on anechoic or heavily damped rooms would clarify what the cue actually is.

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

Quick take: this is a solid first demonstration that a single microphone can extract speech at a queried distance without any speaker embedding, and the same model can estimate speaker distances from the mixture. The idea is clean and the experiments support feasibility, but the evaluation is not as deep as the claim deserves.

What's actually new: the task formulation. Prior distance-based separation (DSS, Rezero) uses static or dynamic thresholds to separate near/far groups; this paper instead conditions directly on a continuous distance in a single-channel TSE framework, with an explicit zero-output rule when no speaker is near the query. That's a genuinely useful reformulation for hearing aids and robot audition, and the distance-estimation byproduct is a nice bonus. The architecture is parameter-efficient (1.25M params) and fuses distance embeddings through both subband and frame dimensions, borrowing sensibly from TF-GridNet/Dual-path RNN.

What the paper does well: it evaluates on four datasets, including real RIRs, and reports means and standard deviations over repeated tests. The comparison against a simple LSTM baseline shows a large improvement, and the performance drop from fixed room (12.47 dB SDR) to random rooms (6.83 dB) to real RIRs (4.79 dB) is honestly reported as a limitation rather than hidden. The distance-estimation result in Fig. 3, where iSDR peaks at the actual speaker distances when scanning the query, is a useful qualitative check.

Soft spots, in proportion: the main concern from the stress-test—that the model might not actually follow the query distance—is only partially valid. The training loss forces the network to produce different outputs for different query distances, because each mixture is trained with both presence and absence cases. A model that ignored dq couldn't satisfy both the presence loss and the zero-output loss simultaneously. Fig. 3 also shows two peaks, not one, which suggests distance conditioning is doing real work. Still, the paper would be much stronger with an explicit wrong-distance control: scramble the query distance and show SDR collapses. That would directly confirm the model is using the cue rather than shortcutting on source salience. As it stands, the evidence is indirect. Also missing are comparisons to strong TSE baselines (e.g., speaker-embedding TSE) or to DSS/Rezero adapted to single-channel continuous-distance settings. The distance estimation results are reported as MAE only, with no error bars or per-room breakdown.

Who this is for: the audio separation community, especially people working on binaural/hearing aid processing and robot audition. It's a niche result but a clean task definition that others can build on.

Recommendation: yes, this deserves peer review. With the right controls and a couple of stronger baselines, it could be a solid conference paper. The current version is a credible feasibility study, and the authors are appropriately modest about its limitations.

Referee Report

3 major / 5 minor

Summary. The paper proposes Distance-Based Single-Channel Target Speech Extraction (TSE), where the only cue for extracting a target speaker is a scalar query distance dq. The architecture consists of a TF-domain encoder, a Distance Query (DQ) block that fuses a learned distance embedding with time-frequency representations via intra-subband and intra-frame BLSTMs, and Temporal/Spectral (TS) blocks for enhancement; training uses an active SDR loss and an inactive SDR loss to handle both presence and absence of speakers near the query distance. The authors evaluate on three simulated RIR datasets (fixed room/fixed mic, fixed room/random mics, random rooms/random mics) and one real RIR dataset (BUT ReverbDB), reporting SDR/SDRi/PESQ improvements and an iSDR metric for absence cases. They also propose a speaker-distance estimation procedure that scans query distances and uses the sum of iSDRs as a detection indicator. The central claim is that this is the first work using only distance information for single-channel TSE, without speaker physiological information.

Significance. If the central claim is fully supported, the paper opens a new cue modality for TSE that avoids biometric/speaker-identity information, which is attractive for privacy-sensitive applications. The model is parameter-efficient (1.25M parameters), and the use of real RIRs in D4, plus the online demonstration, strengthen the practical relevance. The distance-estimation extension is a useful byproduct of the extraction architecture. However, the current evidence does not yet rule out an alternative interpretation in which the model relies partly on source-salience priors rather than on the query distance per se; the missing wrong-distance/no-distance control is load-bearing for the Section I claim. The paper also does not compare with the most closely related distance-conditioned extraction methods, which limits the strength of the effectiveness claim.

major comments (3)

[III-C, Table I, and Fig. 3] The experiments never verify that the extracted speaker is selected by the query distance rather than by source salience. No condition is run with a deliberately wrong dq, nor with the distance embedding replaced by a constant or by random noise. The distance-estimation experiment in Fig. 3 uses iSDR, which is essentially output energy, as the indicator; a model that always extracts the same salient speaker whenever dq lies near any speaker would produce positive SDRi in averaged presence cases and detection peaks at both speaker distances. Please add control experiments (scrambled or wrong query distance, and an ablated model without the distance embedding), and stratify the Table I results by whether the query targets the near versus the far speaker and by whether the query is correct or incorrect. This is necessary to support the Section I statement that the method 'relies solely on distance information.'
[II-A and III-C] The evaluation protocol for the 'presence' condition is underspecified when the query distance falls within rspk of more than one speaker. Equation (3) states that the model should output the sum of speeches in that case, but the paper does not report how often this occurs in the two-speaker mixtures, how the reference signal for SDR is constructed in that situation, or whether such mixtures are excluded from Table I. Since the target speech xs is used in the active loss and in the SDR metric, this ambiguity directly affects the interpretation of every numerical result in Table I. Please clarify the protocol and, if such cases are excluded, state this explicitly.
[III-C, Table II] The only comparison baseline is an LSTM+DEG model described as 'similar to DSS.' Given that the paper claims to be the first distance-only single-channel TSE method, the effectiveness of the proposed architecture should be contextualized against the closest prior distance-conditioned extraction or separation approaches, e.g., adapting the region-customizable sound extraction of [6] or the distance-and-speaker method of [7] to a comparable single-channel, distance-only protocol. Without such comparisons, the 'effectiveness' claim in the abstract is only weakly supported, even though the feasibility claim is supported by the positive SDRi values.

minor comments (5)

[II-B, Eqs. (5)-(7)] The text describing the intra-subband fusion says the distance embedding is 'replicated along the spectral dimension,' but Eq. (5) shows a concatenation that increases the time dimension from T to T+1. Please correct this wording to 'replicated along the temporal dimension' and similarly check the description of the intra-frame fusion for consistency with the actual concatenation axis.
[II-C and Fig. 3] The symbol iSDR is used both as a loss to be minimized (Eq. (9), where lower is better) and as a detection indicator in Fig. 3, where higher values indicate speaker presence. This dual use is confusing and should be clarified with distinct notation or an explicit explanation.
[III-C, Table I] The sentence 'The results are obtained by repeating testing 5 times and expressed as mean ± standard deviation' is unclear about what is repeated: resampling of test mixtures, different random seeds, or something else. Please specify the procedure.
[II-A, Eq. (3)] In the equation 'y dq − → X k xk(t), ks.t. |dk − dq| ≤ rspk', the comma after the summation appears to be a typo; the condition should be written unambiguously as a set comprehension, e.g., summing over {k : |dk − dq| ≤ rspk}.
[III-A, D4] For the real-RIR dataset, please clarify whether the fine-tuning RIRs and test RIRs come from disjoint rooms and whether any test utterances share speakers with the fine-tuning set, as this affects the generalization interpretation of the D4 results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: end-to-end supervised extraction with external data; distance estimation is a use of the trained model, not a fitted prediction.

full rationale

The paper's claimed derivation chain is an empirical training-and-evaluation pipeline, not a formal derivation. The target speech mask is produced by a model whose only conditioning input is the query distance dq; the loss functions (8)-(9) supervise the output against the actual target speech xs when |dk - dq| <= rspk and against zero otherwise. Nothing in the loss injects the test-time result into the model. The four datasets (D1-D4) use simulated RIRs and the external BUT ReverbDB with disjoint fine-tuning/test splits, and the reported SDR/SDRi/PESQ/iSDR numbers are measured on held-out mixtures. The speaker-distance estimation experiment sweeps dq and uses the model's own output-energy indicator to find a peak; this is an application of the trained model's conditioning behavior, not a separate prediction that is fitted to the test labels. No load-bearing step is justified by a self-citation: references [4]-[19] are external prior work, and none of the paper's central claims relies on an author's own uniqueness theorem or ansatz. The absence of a scrambled-query control is a possible validity/control limitation, but it is not a circularity: the model is not defined in terms of the evaluation outcome. Therefore no circular step is present.

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

The central claim rests on a small number of hand-set hyperparameters and on standard room-acoustics and mixture assumptions. No new physical entities are introduced. The main risk is the unvalidated assumption that distance cues are sufficiently discriminative across diverse rooms.

free parameters (3)

rspk (speaker distance range) = 0.5 m for D1-D3, 0.1 m for D4
Defines the range of speaker distances around the query distance that are considered target. Set by hand; affects extraction performance and distance-estimation resolution.
tau (active SDR soft threshold) = 0.001 (eta=30 dB)
Eq. (8); upper-bounds the active SDR loss and is chosen to balance scale accuracy. Not learned.
tau_inactive (inactive loss soft threshold) = 0.01
Eq. (9); soft threshold in the inactive loss that encourages zero output when no speaker is near the query distance. Set by hand.

assumptions (4)

domain assumption The observed single-channel mixture is a linear sum of reverberant speech signals from K speakers.
Eq. (2); assumes no additive noise or nonlinear distortion in the mixture model.
domain assumption Each speaker's signal is generated by convolving anechoic speech with a fixed room impulse response.
Eq. (1); assumes time-invariant RIRs and no source movement during an utterance.
domain assumption The direct-to-reverberation ratio decreases with distance, so distance information is present in the single-channel mixture.
Introduction; this is the physical basis for the method and is not empirically verified in the paper.
domain assumption A learned scalar distance embedding generalizes across rooms and microphone positions.
Section II-B; the distance query blocks assume the embedding trained on simulated data transfers to new acoustic conditions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distance Based Single-Channel Target Speech Extraction." pith.science (2026). https://pith.science/paper/UP5UPMBT

@misc{pith2026241220144,
  author       = {Pith},
  title        = {Pith review of: Distance Based Single-Channel Target Speech Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UP5UPMBT}},
  note         = {Machine review of arXiv:2412.20144}
}

read the original abstract

This paper aims to achieve single-channel target speech extraction (TSE) in enclosures by solely utilizing distance information. This is the first work that utilizes only distance cues without using speaker physiological information for single-channel TSE. Inspired by recent single-channel Distance-based separation and extraction methods, we introduce a novel model that efficiently fuses distance information with time-frequency (TF) bins for TSE. Experimental results in both single-room and multi-room scenarios demonstrate the feasibility and effectiveness of our approach. This method can also be employed to estimate the distances of different speakers in mixed speech. Online demos are available at https://runwushi.github.io/distance-demo-page.

Figures

Figures reproduced from arXiv: 2412.20144 by the authors.

**Figure 2.** (a) Overall structure, (b) Structure of Distance Query block. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Inactive SDR of speech versus different query distance. The red dashed [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID 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

Single-Channel Target Speech Extraction Utilizing Distance and Room Clues
eess.AS 2025-05 conditional novelty 6.0 of 10

Adding room dimensions and reverberation time as inputs to a distance-based target speech extraction model improves SDR by about 1.2 dB on unseen simulated and real rooms.

Reference graph

Works this paper leans on

19 extracted references · 12 canonical work pages · cited by 1 Pith paper

[6]

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
[7]

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

arXiv 2023
[1]

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, May 2023

work page 2023
[2]

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,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2024, pp. 6415–6423, main Track

work page 2024
[3]

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

arXiv 2012
[4]

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), Oct. 2023, pp. 1–5

work page 2023
[5]

Distance-Based Sound Separation,

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

work page 2022
[8]

Multi-channel talker-independent speaker separation through location-based training,

H. Taherian, K. Tan, and D. Wang, “Multi-channel talker-independent speaker separation through location-based training,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 30, pp. 2791– 2800, 2022

work page 2022

Show all 19 references

[9]

Hyperbolic distance-based speech sep- aration,

D. Petermann and M. Kim, “Hyperbolic distance-based speech sep- aration,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 1191–1195

2024
[10]

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
[11]

Das- former: Deep Alternating Spectrogram Transformer For Multi/Single- Channel Speech Separation,

S. Wang, X. Kong, X. Peng, H. Movassagh, V . Prakash, and Y . Lu, “Das- former: Deep Alternating Spectrogram Transformer For Multi/Single- Channel Speech Separation,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , Jun....

2023
[12]

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) . IEEE, 2020, pp. 46–50

2020
[13]

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) , May 2019, pp. 626–630

2019
[14]

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
[15]

What’s all the fuss about free universal sound separation data?

S. Wisdom, H. Erdogan, D. P. 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 Processin...

2021
[16]

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), Apr. 2018, pp. 351–355

2018
[17]

Building and evaluation of a real room impulse response dataset,

I. Sz ¨oke, M. Sk ´acel, L. Mo ˇsner, J. Paliesek, and J. ˇCernock´y, “Building and evaluation of a real room impulse response dataset,” IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 4, pp. 863–876, Aug. 2019

2019
[18]

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
[19]

Adam: A method for stochastic optimization,

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

2014 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [6]

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

[2] [7]

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

arXiv 2023

[3] [1]

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, May 2023

work page 2023

[4] [2]

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,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2024, pp. 6415–6423, main Track

work page 2024

[5] [3]

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

arXiv 2012

[6] [4]

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), Oct. 2023, pp. 1–5

work page 2023

[7] [5]

Distance-Based Sound Separation,

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

work page 2022

[8] [8]

Multi-channel talker-independent speaker separation through location-based training,

H. Taherian, K. Tan, and D. Wang, “Multi-channel talker-independent speaker separation through location-based training,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 30, pp. 2791– 2800, 2022

work page 2022