Pith. sign in

REVIEW 3 major objections 4 minor 51 references

Enhancing Target Speaker Extraction with Explicit Speaker Consistency Modeling

T0 review · 3 major / 4 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 By adding a centroid-based speaker consistency loss and conditional loss suppression to a target speaker extraction system, this paper improves separation quality, extraction accuracy, and speaker similarity on Libri2Mix, with gains that…

desk verdict A modest, plausibly real training-objective gain, but the joint-training setup hides a cross-encoder alignment that the paper never discusses, and the SOTA claim outruns Table 3. read the letter →

arxiv 2507.09510 v3 pith:VLBORAL4 submitted 2025-07-13 cs.SD eess.AS

Shu Wu , Anbin Qi , Yanzhang Xie , Xiang Xie This is my paper
classification cs.SDeess.AS
keywords targetspeakerextractionconsistencycentroid-basedlossconditionalsuppressionembeddingconfusionSI-SDRcocktailpartyproblem
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

Target speaker extraction pulls one voice out of a mixture using a reference recording of that voice. This paper argues that previous systems focus on making the reference embedding better but ignore whether the extracted voice still sounds like the same speaker. It proposes a centroid-based speaker consistency loss that pushes the embedding of the extracted speech toward the average embedding of the target speaker, plus a conditional suppression rule that stops applying that loss once similarity is high. On the Libri2Mix dataset this improves separation quality, extraction accuracy, and speaker similarity for two speaker encoders, and the improvement carries over to two other separator architectures.

What carries the argument

The load-bearing object is the speaker centroid $e^C_i$, the mean of utterance-level embeddings for a speaker, used in a softmax cosine loss that pulls the extracted speech's embedding toward the target speaker's centroid while pushing it away from other speakers' centroids. The second piece is conditional loss suppression, a gate $f_C(x)=x$ if $\mathrm{SECS}\le\omega$ else $0$, with $\omega$ decreasing linearly from 1.0 to 0.8, which turns off the consistency loss once the model has achieved high speaker similarity so the separator is not over-constrained.

What would settle it

Retrain the proposed system with many random seeds and check whether the SI-SDR and accuracy gains over the baseline persist; if they vanish or reverse, the result is training noise. Separately, measure whether the consistency loss raises cosine similarity to the correct speaker centroid without also raising similarity to a wrong speaker's centroid, because simultaneous inflation would show the loss improves similarity without true identity preservation.

Watch

Extended reading notes

Core claim

The central claim is that adding an explicit speaker-consistency objective to a target speaker extraction system improves both extraction quality and identity preservation. The loss is $L_{C-SC} = -\log \frac{\exp(\cos(\hat{e}_s, e^C_{i_T}))}{\sum_{i=1}^N \exp(\cos(\hat{e}_s, e^C_i))}$, where $\hat{e}_s$ is the speaker embedding of the extracted speech and $e^C_i$ is the centroid (average utterance-level embedding) of speaker $i$; this is combined with SI-SDR loss and, when the encoder is trainable, speaker classification loss. The authors report that this loss improves SI-SDR, extraction accuracy, and speaker similarity over the baseline in all four encoder/training configurations, and that conditional loss suppression (zeroing the consistency term once cosine similarity exceeds a threshold that decays from 1.0 to 0.8) adds further SI-SDR and accuracy gains.

Load-bearing premise

The load-bearing premise is that two different utterances by the same speaker should map to nearby speaker embeddings, so maximizing cosine similarity between the extracted speech embedding and the speaker's centroid is a safe objective; if the speaker encoder is sensitive to recording conditions, content, or duration rather than speaker identity alone, this loss could push the separator toward artifacts, and the reported single-run gains could be training noise.

Editorial extensions

If this is right

  • The combined method (centroid consistency loss plus conditional loss suppression) improves SI-SDR, extraction accuracy, and speaker similarity over the BSRNN baseline in all four encoder/training configurations, with joint-trained ECAPA-TDNN reaching 96.70% extraction accuracy.
  • Conditional loss suppression adds SI-SDR and accuracy on top of the consistency loss in every configuration, at a small cost in speaker similarity relative to using the consistency loss alone.
  • The same training objective transfers to DPCCN and TF-GridNet backbones, raising SI-SDR and accuracy on both.
  • Models trained with the proposed method also outperform the compared target speaker extraction systems on Libri2Mix in SI-SDRi and accuracy, including approaches that use self-supervised speech models.

Reading between the lines

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

  • Because the centroid bank is computed offline from the training speaker set, a natural testable extension would update centroids dynamically from enrollment clips at inference time, which would let the loss apply to speakers never seen in training.
  • The reported speaker similarity metric is computed with a different encoder than the one used in the training loss, so part of the measured gain could reflect objective mismatch rather than genuine identity preservation; a same/different speaker verification test on the extracted audio would settle this.
  • If the mechanism is genuine identity anchoring, the same loss should also help under noisy or reverberant conditions where enrollment and target differ acoustically; that prediction is not tested in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a method to improve target speaker extraction (TSE) by explicitly modeling speaker consistency between the enrollment speech and the extracted speech. The main contribution is a centroid-based speaker consistency loss (LC-SC) that encourages the embedding of the extracted speech to lie near the target speaker's centroid, plus a conditional loss suppression (CLS) strategy that disables this loss once the enrollment-vs-extracted cosine similarity exceeds a threshold. The method is evaluated on Libri2Mix using BSRNN as the backbone with ECAPA-TDNN and ResNet34 speaker encoders in both pretrained and joint training modes. The reported results show consistent improvements in SI-SDR, extraction accuracy, PESQ, STOI, and speaker similarity, along with cross-backbone (DPCCN, TF-GridNet) and out-of-domain (Aishell2Mix) generalization. The authors claim state-of-the-art performance.

Significance. The idea of encouraging speaker consistency during TSE training is well motivated, and the use of an independent encoder (ECAPA2) for the similarity metric is a notable strength that avoids circularity in evaluation. The systematic ablations and generalization tests across backbones and datasets are valuable. If the reported gains are reproducible, the proposed loss is a simple and potentially useful add-on to existing TSE systems. However, the central claims rest on single-run experiments with small margins, and the joint-training regime is under-specified with respect to how the speaker centroids are obtained, which limits the interpretability of the joint-training results.

major comments (3)
  1. [§3.2, Eqs. (4)–(5)] In joint training mode, the paper states that ECAPA-TDNN and ResNet34 are 'trained from scratch with the speech separator model,' but it also says these pretrained models are used 'for obtaining eU_i,k in Section 2.2.3.' This implies the centroids eC_i in Eq. (4) are computed once with the frozen pretrained encoder, while the embeddings e_hat_s in Eq. (5) come from the randomly initialized, jointly trained encoder. The loss then pushes the joint encoder's outputs toward a fixed, external embedding space, which is a cross-encoder distillation that is never described or analyzed. If, instead, the centroids are recomputed during joint training, that procedure is not stated. The paper must clarify this and, if the centroids are fixed, provide an analysis of the compatibility of the two embedding spaces, or re-estimate the centroids during joint training. As written, the improvements reported for the joint-training rows in Table 1 are not fully interpretable.
  2. [Tables 1, 2, 4, 5] All experiments are single-run with no error bars, no multiple seeds, and no significance tests. Given that the reported gains are small (e.g., SI-SDR improvements of 0.43–0.60 dB, accuracy gains of about 1–2 percentage points, and Sim. gains of about 1–3 percentage points in Table 1), the possibility that these differences arise from training noise cannot be excluded. Please provide mean and standard deviation over at least three random seeds for the main comparisons, or a significance test, to support the claim that the proposed losses reliably improve performance.
  3. [Table 3, abstract, §4.3] The comparison with prior methods combines results cited from different papers without specifying their training data, evaluation protocols, or whether the 'accuracy' metric is defined identically across all systems. The margin over the best prior systems is small (e.g., 14.75 vs. 14.65 dB SI-SDRi for SSL-TD-SpeakerBeam), and accuracy values are missing for several baselines. The 'state-of-the-art' claim in the abstract and Section 4.3 is therefore not substantiated unless a controlled comparison under the same data and protocol is provided.
minor comments (4)
  1. [§2.2.1, Eq. (1)] The notation for the embedding of the estimated speech is inconsistent ('e_hat_s' and 'es'), and the text mentions 'ˆer' without definition. Please unify the notation throughout.
  2. [Tables 1, 2, 4, 5] Several table cells contain duplicated numbers (e.g., '86.9286.9286.92' and '14.2914.2914.29'), likely a formatting artifact. Please correct these.
  3. [§2.3.2, Eq. (7)] The conditional loss suppression function f_C gates the centroid-based loss based on the SECS value, but the relationship between the gating condition (which uses the enrollment/extracted cosine similarity) and the gated term (which is the centroid loss) is not explained. Please clarify whether the gating applies only to LC-SC or to all consistency losses, and why using SECS as the condition is appropriate.
  4. [Table 5] The text 'train the models on the Libri2Mix-train-100 and VoxCeleb1 datasets' is ambiguous; please state explicitly that these are two separate training runs, and specify the number of speakers in each training set and whether the evaluation sets share speakers with the training sets.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the proposed losses are supervised training objectives, and the reported speaker-similarity metric uses an independent ECAPA2 encoder.

full rationale

The paper's central claim is empirical: adding the centroid-based speaker consistency loss LC-SC and conditional loss suppression CLS to a BSRNN TSE system improves SI-SDR, extraction accuracy, and speaker similarity on Libri2Mix. LC-SC (Eq. 5) is a supervised training loss that maximizes the cosine similarity between the estimated-speech embedding and pretrained speaker centroids; it is an optimization objective, not a prediction derived from the data it evaluates. No fitted parameter is renamed as a prediction: the hyperparameters beta, lambda, and omega are fixed constants reported in Sections 2.3.1 and 2.3.2, and the model is evaluated on held-out test speakers with standard metrics (SI-SDR, SDR, PESQ, STOI, accuracy). Critically, the reported Sim. metric is computed with the ZS-TTS-Evaluation toolkit using ECAPA2 as the speaker encoder, which is independent of the WeSpeaker ECAPA-TDNN and ResNet34 encoders used to compute the centroids and the training loss; therefore the reported similarity improvement is not the same function being optimized by construction. The baseline comparisons are BSRNN, DPCCN, and TF-GridNet without the proposed losses, and Table 3 compares against results cited from prior papers; there are no load-bearing self-citations that reduce the central claim to an author's prior assertion. The joint-training configuration uses centroids from a frozen VoxCeleb2-pretrained encoder while the jointly trained encoder is randomly initialized, which is an internal design detail that could affect interpretation of the joint-mode gains, but it does not make the result true by definition: the loss could fail to improve SI-SDR or ECAPA2 similarity, and the paper reports consistent gains. This is a methodological under-specification, not circularity. The derivation chain is therefore self-contained with respect to the stated objectives and external evaluation metrics, and no step reduces to its own inputs.

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

The central claim rests on the assumption that speaker embeddings are faithful, content-invariant representations of identity, on standard supervised separation data, and on three manually chosen hyperparameters (beta, lambda, omega). No new physical or theoretical entities are introduced; the speaker centroid is a mean of embeddings from an existing pretrained encoder.

free parameters (3)
  • Consistency loss weight lambda = 0.1
    Weight of LC-SC in Eq. (6) and Eq. (8); set to 0.1 with no reported sensitivity analysis.
  • CE loss weight beta = 0.1 in joint training, 0 when frozen
    Weight of speaker classification loss in Eq. (6); chosen manually and not ablated.
  • SECS suppression threshold omega = linearly from 1.0 to 0.8
    Threshold in Eq. (7) that switches off the consistency loss; the schedule is hand-set with no sensitivity study.
assumptions (3)
  • domain assumption Two utterances by the same speaker map to close speaker embeddings under the speaker encoder, regardless of content differences.
    Justifies the speaker consistency loss in Section 2.2.2; if the encoder is content-sensitive, the loss may not correspond to true speaker identity.
  • domain assumption Cosine similarity between speaker embeddings is a valid measure of speaker identity consistency.
    Used in Eq. (1) to define SECS and in the external Sim. metric via ECAPA2; the paper does not validate this against human judgments in the TSE setting.
  • domain assumption The WeSpeaker pretrained encoders, trained on VoxCeleb2, generalize to Libri2Mix and Aishell2Mix speakers for the purpose of computing centroids and consistency losses.
    Centroids and enrollment embeddings in Section 3.2 rely on this transfer; no speaker verification accuracy on the target datasets is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Target Speaker Extraction with Explicit Speaker Consistency Modeling." pith.science (2026). https://pith.science/paper/VLBORAL4

@misc{pith2026250709510,
  author       = {Pith},
  title        = {Pith review of: Enhancing Target Speaker Extraction with Explicit Speaker Consistency Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLBORAL4}},
  note         = {Machine review of arXiv:2507.09510}
}
read the original abstract

Target Speaker Extraction (TSE) uses a reference cue to extract the target speech from a mixture. In TSE systems relying on audio cues, the speaker embedding from the enrolled speech is crucial to performance. However, these embeddings may suffer from speaker identity confusion. Unlike previous studies that focus on improving speaker embedding extraction, we improve TSE performance from the perspective of speaker consistency. In this paper, we propose a speaker consistency-aware target speaker extraction method that incorporates a centroid-based speaker consistency loss. This approach enhances TSE performance by ensuring speaker consistency between the enrolled and extracted speech. In addition, we integrate conditional loss suppression into the training process. The experimental results validate the effectiveness of our proposed methods in advancing the TSE performance. A speech demo is available online:https://sc-tse.netlify.app/

Figures

Figures reproduced from arXiv: 2507.09510 by the authors.

Figure 1
Figure 1. The architecture of the proposed speaker consistency-aware target speaker extraction model 2.2. Centroid-based speaker consistency loss 2.2.1. Speaker consistency in TSE Speaker similarity is commonly used to assess the degree of similarity between the speech of different speakers. In the field of Text-To-Speech (TTS) study [22], speaker encoder cosine similarity (SECS) is utilized to evaluate the speaker consistenc… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 47 canonical work pages

  1. [1]

    The chal- lenge of isolating the target speech while ignoring other inter- ferences is known as the cocktail party problem [1]

    Introduction In real-world scenarios, we navigate complex acoustic environ- ments filled with overlapping speech and diverse background noise, such as music or machine-generated sounds. The chal- lenge of isolating the target speech while ignoring other inter- ferences is known as the cocktail party problem [1]. Humans achieve this through selective atten...

    work page 2017

  2. [2]

    Proposed methods 2.1. Model architecture overview We adopt a standard speaker extraction pipeline (Figure 1), comprising a speaker encoder, a separation module, and the pro- posed speaker consistency loss module. The Band-Split RNN (BSRNN) [21] serves as the separator backbone. The speaker encoder is either pre-trained or jointly trained with the separa- ...

    work page Pith review arXiv 2025

  3. [3]

    The proposed methods enhance overall performance, achieving state-of-the-art results

    We conduct a comprehensive evaluation to validate the effec- tiveness of the proposed methods for target speaker extraction. The proposed methods enhance overall performance, achieving state-of-the-art results. The ablation study further confirmed the effectiveness of the methods. Meanwhile, experiments on dif- ferent backbones and out-of-domain datasets ...

    work page

  4. [4]

    Dataset We use the clean Libri2Mix [26] dataset with two-speaker mix- tures

    Experiments 3.1. Dataset We use the clean Libri2Mix [26] dataset with two-speaker mix- tures. TSE models are trained on the 100-hour subset (13900 utterances, 251 speakers), with validation and test sets contain- ing 3,000 utterances from 40 non-overlapping speakers. Each sample is used twice for different speaker extractions by alter- nating enrollment. ...

    work page

  5. [5]

    Conclusion and future work In this paper, we propose a speaker consistency-aware target speaker extraction that integrates centroid-based speaker con- sistency loss and conditional loss suppression. Experiments demonstrate the effectiveness of our approach, showing sub- stantial improvements not only in the speaker consistency be- tween the extracted and ...

    work page

  6. [6]

    Comparative studies with proposed methods Table 1 presents the performance of BSRNN with proposed methods on Libri2Mix

    Results 4.1. Comparative studies with proposed methods Table 1 presents the performance of BSRNN with proposed methods on Libri2Mix. Four comparative studies were con- ducted using ECAPA-TDNN and ResNet34 models under pre- trained or joint training modes, without utilizing the speaker centroid. In each case, the BSRNN model combined with a speaker encoder...

    work page

  7. [7]

    and other metrics in all scenarios

    The centroid-based speaker consistency loss improve both Sim. and other metrics in all scenarios. With pretrained ECAPA-TDNN, SI-SDR improved by 0.51 dB, accuracy by 1.02%, and Sim. by 2.64%, alongside consistent gains across metrics. Joint training show similar trends, with SI- SDR improving by 0.60 dB, accuracy by 1.92%, and Sim. by 1.36%

    work page

  8. [8]

    For ResNet34 in pretrained mode, SI-SDR improved by 0.43 dB, accuracy by 1.56%, and Sim

    The centroid-based speaker consistency loss demonstrates strong generalization across speaker encoders. For ResNet34 in pretrained mode, SI-SDR improved by 0.43 dB, accuracy by 1.56%, and Sim. by 0.97%. Similar gains are observed in joint training, with comparable increases in all metrics

    work page

Show all 51 references
  1. [9]

    However, a slight decrease in the Sim

    The CLS strategy enhances TSE system performance, im- proving SI-SDR, accuracy, and other metrics in all scenarios. However, a slight decrease in the Sim. metric was observed, though CLS still outperformed the baseline. This may re- sult from the suppressive effect of CLS on c...

  2. [10]

    Overall, by combining centroid-based speaker consistency loss with the CLS strategy, our proposed methods lead to improvements in the TSE system’s performance. 4.2. Ablation study Table 2 illustrates the impact of integrating the speaker cen- troid and CLS into the speaker con...

    1915

  3. [11]

    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,” in Interspeech 2020, Oct 2020. [Online]. Available: http://dx.doi.org/10.21437/interspeech.2020-1397

    2020 doi

  4. [12]

    Some further experiments upon the recognition of speech, with one and with two ears,

    E. C. Cherry and W. Taylor, “Some further experiments upon the recognition of speech, with one and with two ears,”The Journal of the acoustical society of America, vol. 26, no. 4, pp. 554–559, 1954

    1954

  5. [13]

    The cocktail-party problem revisited: early process- ing and selection of multi-talker speech,

    A. W. Bronkhorst, “The cocktail-party problem revisited: early process- ing and selection of multi-talker speech,” Attention, Perception, & Psy- chophysics, vol. 77, no. 5, pp. 1465–1487, 2015

    2015

  6. [14]

    Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,

    K. ˇZmol´ıkov´a, M. Delcroix, K. Kinoshita, T. Ochiai, T. Nakatani, L. Burget, and J. ˇCernock´y, “Speakerbeam: Speaker aware neural network for target speaker extraction in speech mixtures,” IEEE Journal of Selected Topics in Signal Processing, vol. 13, no. 4, pp. 800–814, 2019

    2019

  7. [15]

    Adaptive blind audio source extraction supervised by dom- inant speaker identification using x-vectors,

    J. Jansk `y, J. M ´alek, J. ˇCmejla, T. Kounovsk `y, Z. Koldovsk `y, and J. ˇZd’´ansk`y, “Adaptive blind audio source extraction supervised by dom- inant speaker identification using x-vectors,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal ...

    2020

  8. [16]

    Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation,

    A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K. Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein, “Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation,” ACM Transactions on Graphics , p. 1–11, Aug 2018. [Online]. Available:...

    2018

  9. [17]

    Audio-visual sound separation via hidden markov models,

    J. Hershey and M. Casey, “Audio-visual sound separation via hidden markov models,” Advances in Neural Information Processing Systems , vol. 14, 2001

    2001

  10. [18]

    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 , Sep 2019. [Online]. Available: http://dx.doi.org/10.21437/interspeech.2019-2266

    2019 doi

  11. [19]

    V oice filter: Few- shot text-to-speech speaker adaptation using voice conversion as a post- processing module,

    A. Gabry ´s, G. Huybrechts, M. S. Ribeiro, C.-M. Chien, J. Roth, G. Co- mini, R. Barra-Chicote, B. Perz, and J. Lorenzo-Trueba, “V oice filter: Few- shot text-to-speech speaker adaptation using voice conversion as a post- processing module,” in ICASSP 2022-2022 IEEE Internatio...

    2022

  12. [20]

    Single-channel speech extraction using speaker inventory and attention network,

    X. Xiao, Z. Chen, T. Yoshioka, H. Erdogan, C. Liu, D. Dimitriadis, J. Droppo, and Y . Gong, “Single-channel speech extraction using speaker inventory and attention network,” in ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)...

    2019

  13. [21]

    Target confusion in end- to-end speaker extraction: Analysis and approaches,

    Z. Zhao, D. Yang, R. Gu, H. Zhang, and Y . Zou, “Target confusion in end- to-end speaker extraction: Analysis and approaches,” in Proc. Interspeech 2022, 2022, pp. 5333–5337

    2022

  14. [22]

    Sc-glowtts: an efficient zero-shot multi-speaker text-to-speech model,

    E. Casanova, C. Shulby, E. G ¨olge, N. M. M ¨uller, F. S. d. Oliveira, A. Candido Jr., A. d. S. Soares, S. M. Aluisio, and M. A. Ponti, “Sc-glowtts: an efficient zero-shot multi-speaker text-to-speech model,” in Interspeech 2021 , Aug 2021. [Online]. Available: http: //dx.doi....

    2021 doi

  15. [23]

    X-sepformer: End-to-end speaker extraction network with explicit optimization on speaker confusion,

    K. Liu, Z. Du, X. Wan, and H. Zhou, “X-sepformer: End-to-end speaker extraction network with explicit optimization on speaker confusion,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  16. [24]

    Speaker ex- traction with detection of presence and absence of target speakers,

    K. Zhang, M. Borsdorf, Z. Pan, H. Li, Y . Wei, and Y . Wang, “Speaker ex- traction with detection of presence and absence of target speakers,” inProc. Interspeech 2023, 2023, pp. 3714–3718

    2023

  17. [25]

    On the effectiveness of enrollment speech augmentation for target speaker extraction,

    J. Li, K. Zhang, S. Wang, H. Li, M.-W. Mak, and K. A. Lee, “On the effectiveness of enrollment speech augmentation for target speaker extraction,” 2024 IEEE Spoken Language Technology Workshop (SLT) , pp. 325–332, 2024. [Online]. Available: https://api.semanticscholar.org/ Cor...

    2024

  18. [26]

    Selective hubert: Self- supervised pre-training for target speaker in clean and mixture speech,

    J. Lin, M. Ge, W. Wang, H. Li, and M. Feng, “Selective hubert: Self- supervised pre-training for target speaker in clean and mixture speech,” IEEE Signal Processing Letters, 2024

    2024

  19. [28]

    Contrastive learning for target speaker extraction with attention-based fusion,

    X. Li, R. Liu, H. Huang, and Q. Wu, “Contrastive learning for target speaker extraction with attention-based fusion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 178–188, 2024

    2024

  20. [29]

    Real-time personalised speech enhancement transformers with dynamic cross-attended speaker representations,

    S. Zhang, M. Chadwick, A. Ramos, T. Parcollet, R. van Dalen, and S. Bhat- tacharya, “Real-time personalised speech enhancement transformers with dynamic cross-attended speaker representations,” inProc. Interspeech 2023, 2023, pp. 804–808

    2023

  21. [30]

    Smma-net: An audio clue- based target speaker extraction network with spectrogram matching and mu- tual attention,

    Y . Hu, H. Xu, Z. Guo, H. Huang, and L. He, “Smma-net: An audio clue- based target speaker extraction network with spectrogram matching and mu- tual attention,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, p...

    2024

  22. [31]

    Target speaker extraction by di- rectly exploiting contextual information in the time-frequency domain,

    X. Yang, C. Bao, J. Zhou, and X. Chen, “Target speaker extraction by di- rectly exploiting contextual information in the time-frequency domain,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 476–10 480

    2024

  23. [32]

    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 Processing , vol. 31, pp. 1893–1901, 2023

    1901

  24. [33]

    An algorithm for intelligibility prediction of time–frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An algorithm for intelligibility prediction of time–frequency weighted noisy speech,” IEEE Transactions on audio, speech, and language processing, vol. 19, no. 7, pp. 2125–2136, 2011

    2011

  25. [34]

    In defence of metric learning for speaker recognition,

    J. S. Chung, J. Huh, S. Mun, M. Lee, H.-S. Heo, S. Choe, C. Ham, S.-Y . Jung, B.-J. Lee, and I. Han, “In defence of metric learning for speaker recognition,” in Interspeech, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:214667019

    2020

  26. [35]

    Centroid estimation with transformer-based speaker embedder for robust target speaker extraction,

    W.-H. Heo, J. Maeng, Y . Kang, and N. Cho, “Centroid estimation with transformer-based speaker embedder for robust target speaker extraction,” in Proc. Interspeech 2024, 2024, pp. 4333–4337

    2024

  27. [36]

    Sdr–half-baked or well done?

    J. Le 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). IEEE, 2019, pp. 626– 630

    2019

  28. [37]

    Lib- rimix: An open-source dataset for generalizable speech separation,

    J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vincent, “Lib- rimix: An open-source dataset for generalizable speech separation,”Le Cen- tre pour la Communication Scientifique Directe - HAL - memSIC,Le Centre pour la Communication Scientifique Directe - HAL - memS...

    2020

  29. [38]

    But system description to voxceleb speaker recognition challenge 2019,

    H. Zeinali, S. Wang, A. Silnova, P. Mat ˇejka, and O. Plchot, “But system description to voxceleb speaker recognition challenge 2019,” in Proceedings of The VoxCeleb Challange Workshop 2019 , 2019, pp. 1–4. [Online]. Available: https://www.fit.vut.cz/research/publication/12224

    2019

  30. [39]

    Ecapa-tdnn : Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn : Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” in Interspeech 2020, Oct 2020. [Online]. Available: http://dx.doi.org/10.21437/interspeech.2020-2650

    2020 doi

  31. [40]

    Wespeaker: A research and production oriented speaker embed- ding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embed- ding learning toolkit,” in ICASSP 2023 - 2023 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 202...

    2023

  32. [41]

    V oxceleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,” in Interspeech 2018, Aug 2018. [Online]. Available: http://dx.doi.org/10.21437/interspeech.2018-1929

    2018 doi

  33. [42]

    Wesep: A scalable and flexible toolkit towards generalizable target speaker extraction,

    S. Wang, K. Zhang, S. Lin, J. Li, X. Wang, M. Ge, J. Yu, Y . Qian, and H. Li, “Wesep: A scalable and flexible toolkit towards generalizable target speaker extraction,” 2024. [Online]. Available: https://arxiv.org/abs/2409.15799

    2024 arXiv

  34. [43]

    Perceptual evaluation of speech quality (pesq): An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs,

    I.-T. Recommendation, “Perceptual evaluation of speech quality (pesq): An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs,” Rec. ITU-T P . 862, 2001

    2001

  35. [45]

    Xtts: a massively mul- tilingual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi, and J. Weber, “Xtts: a massively mul- tilingual zero-shot text-to-speech model,” in Interspeech 2024, 2024, pp. 4978–4982

    2024

  36. [46]

    Ecapa2: A hybrid neural network archi- tecture and training strategy for robust speaker embeddings,

    J. Thienpondt and K. Demuynck, “Ecapa2: A hybrid neural network archi- tecture and training strategy for robust speaker embeddings,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2023, pp. 1–8

    2023

  37. [47]

    Multi-level speaker representation for target speaker extraction,

    K. Zhang, J. Li, S. Wang, Y . Wei, Y . Wang, Y . Wang, and H. Li, “Multi-level speaker representation for target speaker extraction,” 2024. [Online]. Available: https://arxiv.org/abs/2410.16059

    2024 arXiv

  38. [48]

    Dpccn: Densely-connected pyramid complex convolutional network for robust speech separation and extraction,

    J. Han and Y . Long, “Dpccn: Densely-connected pyramid complex convolutional network for robust speech separation and extraction,” ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 7292–7296, 2021. [Online]. Available: h...

    2022

  39. [49]

    Mc-spex: Towards effective speaker extraction with multi-scale interfu- sion and conditional speaker modulation,

    J. Chen, W. Rao, Z. Wang, J. Lin, Y . Ju, S. He, Y . Wang, and Z. Wu, “Mc-spex: Towards effective speaker extraction with multi-scale interfu- sion and conditional speaker modulation,” in Proc. Interspeech 2023, 2023, pp. 4034–4038

    2023

  40. [50]

    Tar- get speech extraction with pre-trained self-supervised learning models,

    J. Peng, M. Delcroix, T. Ochiai, O. Plchot, S. Araki, and J. ˇCernock`y, “Tar- get speech extraction with pre-trained self-supervised learning models,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 421–10 425

    2024

  41. [51]

    Tf- gridnet: Making time-frequency domain models great again for monaural speaker separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watanabe, “Tf- gridnet: Making time-frequency domain models great again for monaural speaker separation,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  42. [52]

    V oxceleb: A large-scale speaker identification dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: A large-scale speaker identification dataset,” in Proceedings of Interspeech , 2017, pp. 2616–2620

    2017

  43. [128]

    The speaker separator uses a 192-dimensional bidirec- tional LSTM, repeated 6 times in the BSRNN. ResNet34 [27] and ECAPA-TDNN [28] from the WeSpeaker toolkit 2 [29], pretrained on 5,994 speakers from V oxCeleb2 [30], serve as speaker encoders in pretrain mode and for obtainin...

Pith tools

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