Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Target speaker extraction works better when cue modules and separators are decoupled, so any mix of enrollment, spatial, visual, or text cues can train under one pipeline—even when some cues are missing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 00:56 UTC pith:4R2XN7BP

load-bearing objection Solid modular TSE toolkit paper with clean same-backbone cue comparisons; incremental over prior WeSep, useful for the subfield, soft mainly on zero-fill missingness and promised-not-shipped code. the 3 major comments →

arxiv 2607.27436 v1 pith:4R2XN7BP submitted 2026-07-29 eess.AS

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

Ke Zhang , Xiaoyang Yu , Haoyu Li , Shuai Wang , Shuhan Zhang , Haizhou Li This is my paper
classification eess.AS
keywords target speaker extractionmulti-modal cue modelingmodular architecturespeech separationcocktail-party problemheterogeneous cue conditioningmissing-cue robustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Target speaker extraction (TSE) pulls one desired voice out of overlapping speech using auxiliary cues such as enrollment audio, direction of arrival, lip video, or keywords. Most systems hard-wire a single cue type into the architecture, so they break or need redesign when that cue is unavailable or when several cues should be combined. This paper argues that TSE should be treated as heterogeneous cue-conditioned learning: cue frontends and separator backbones stay separate, talk through fixed interfaces, and are composed by configuration. That design lets researchers swap cue features inside a modality, stack cues across modalities, and train on batches where different samples have different cue subsets—including samples with missing cues represented as zeros—without rebuilding the network. Experiments on enrollment, spatial, visual, and textual cues show clear modality-dependent gains, complementary multi-cue improvements, and stable performance when cues are randomly absent. The practical stake is selective listening systems that can adapt as real-world cue availability changes.

Core claim

WeSep establishes that target speaker extraction can be reformulated and implemented as heterogeneous cue-conditioned learning: by decoupling modality-specific cue frontends from separator backbones through standardized injection interfaces and a sample-level heterogeneous data pipeline, single-cue, multi-cue, and missing-cue settings share one optimization framework, and experiments confirm modality-dependent extraction quality plus stable training when enrollment and spatial cues are partially or jointly absent.

What carries the argument

Modular Top Model composition: config-instantiated cue frontends (speaker, spatial, visual, textual) inject standardized representations into interchangeable separator backbones (e.g., BSRNN) at chosen fusion points, while the data layer indexes mix–target pairs against modality repositories and masks or zero-fills unavailable cues per sample.

Load-bearing premise

Treating missing cues as simple zero placeholders in a shared network is enough to learn real dynamic cue failure, without special missingness models or architecture changes.

What would settle it

Train the same multi-cue model with 30% random cue dropout via zero-fill, then test on structured real failures (correlated visual occlusion, biased DOA error, enrollment mismatch); if performance collapses relative to always-present cues or to a model with explicit missingness handling, the heterogeneous-training claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • New cue types or feature variants can be added by writing a frontend and config, without redesigning the separator.
  • Intra-modal feature combinations (e.g., spectral plus contextual speaker cues) and cross-modal stacks (speaker plus spatial) become controlled ablations under one backbone and loss.
  • A single trained model can serve scenarios where enrollment, location, video, or keywords appear in different subsets per utterance.
  • Causal, low-latency deployments remain compatible with multi-feature conditioning under the same modular interfaces.
  • Public toolkit release enables reproducible comparison of cue structure and availability effects across labs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If zero-fill works mainly because absence is random and independent, production systems may still need uncertainty-aware fusion when cue failures are correlated (e.g., dark room kills video and confuses DOA together).
  • The same decoupling pattern could extend to other conditioned audio tasks—target sound extraction, personalized enhancement—where side information is intermittent.
  • Keyword/text cues matching or beating strong audio enrollment under the same backbone suggests language-level conditioning may reduce enrollment friction in consumer devices.
  • Preserving fine-grained T-F spatial features outperforming compressed spatial embeddings hints that interface design should prefer high-resolution physical cues when the backbone can consume them.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents WeSep, a modular toolkit that reformulates target speaker extraction (TSE) as heterogeneous cue-conditioned learning. Cue frontends (enrollment, spatial, visual, textual) are decoupled from separator backbones (BSRNN, NBC2, etc.) via standardized injection interfaces and a configuration-driven Top Model, so single-cue, multi-cue, and sample-level missing-cue settings share one data and optimization pipeline. Experiments with a fixed BSRNN backbone and schedule report modality-dependent SI-SDRi gains (spectral/frame speaker features over utterance embeddings; handcrafted T-F spatial features over compressed embeddings; keyword text competitive with strong audio cues; visual cues improved by a stronger backbone) and non-collapse under random cue absence with zero-fill placeholders (Tables 1–7). A public toolkit release is promised.

Significance. If the modular interfaces and heterogeneous pipeline behave as claimed, WeSep is a useful experimental substrate for the TSE community: it lowers the cost of comparing cue formulations and combinations under matched separators and training, and it directly addresses dynamic cue availability, which most cue-specific systems ignore. Strengths include controlled same-backbone ablations, coverage of four cue families, a causal/streaming export note, and an intended open-source release. The contribution is primarily architectural and systems-oriented rather than a new separation algorithm; that is appropriate for a framework paper and still valuable if the interfaces are genuinely reusable.

major comments (3)
  1. [Section 4.6.2, Table 7] Section 4.6.2 and Table 7 support stability under random cue absence, but the only missingness mechanism is zero-valued placeholders. This is load-bearing for the claim that heterogeneous availability needs no architectural change. The manuscript should either (i) ablate alternatives (e.g., learned missing tokens, explicit availability masks, or cue-dropout schedules) or (ii) clearly bound the claim to random absence and discuss failure modes when missingness is structured (occlusion, biased DOA error, enrollment mismatch). Without that, transfer to real dynamic reliability remains an assumption rather than a demonstrated result.
  2. [Section 4.6.1, Table 6; Sections 4.4–4.5] Composable multi-cue integration is central (Abstract; Section 3.4), yet joint conditioning is shown only for enrollment+spatial (Table 6). Visual (Table 5) and textual (Table 4) results are single-cue on different corpora. At least one controlled multi-cue run that includes visual and/or textual cues with the same backbone—or an explicit statement that full cross-modal composition was not evaluated—would better support the “cue-composable” claim and the unified-optimization narrative.
  3. [Abstract; Section 5; Table 7] Heterogeneous training is demonstrated only for speaker and spatial cues on the multi-channel set (Table 7), while enrollment/text use Libri2Mix and vision uses VoxCeleb2-mix. That split is reasonable for modality-specific tests, but the paper’s strongest framing—“stable optimization under heterogeneous cue availability” across enrollment, spatial, visual, and textual cues—overreaches the single two-cue missingness protocol. Please narrow the abstract/conclusion wording to match the actual heterogeneous protocol, or add a multi-repository batch experiment that mixes more than two cue types under partial absence.
minor comments (5)
  1. [Throughout] Several proper names are split by spaces in the compiled text (e.g., “ClearerV oice”, “V oxCeleb2”, “LExt” context is fine but check “gpuRIR”/author names). Please proof the PDF for encoding/line-break artifacts.
  2. [Figure 2] Figure 2 is dense; a short caption callout listing which injection points were used for each cue in the default BSRNN setup would help readers reproduce the Top Model configs without reading the code.
  3. [Table 1] Table 1 “Accuracy” (SI-SDRi > 1 dB) is useful but unconventional; one sentence defining why 1 dB is the success threshold would avoid ambiguity.
  4. [Section 1] Related multi-cue TSE work (e.g., multimodal attention fusion [17], text-guided extraction [18]) is cited; a brief explicit contrast of WeSep’s shared pipeline versus independent multi-model toolkits (ClearerVoice) in one paragraph would sharpen novelty without lengthening much.
  5. [Section 4.2.1] Causal results (Table 2) report theoretical 32 ms latency and ONNX export; stating real-time factor or measured end-to-end latency on a reference device would strengthen the deployment claim.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with external SI-SDRi benchmarks, not a self-forcing derivation.

full rationale

WeSep is an architectural/toolkit paper that reformulates TSE as configurable cue-conditioned learning and reports SI-SDRi on held-out mixtures under fixed backbones and schedules. There is no first-principles derivation chain, uniqueness theorem, or fitted parameter re-presented as a prediction. Self-citations (prior WeSep, USEF, TF-Map, MuSE, DAE-TSE, etc.) supply modular building blocks and baselines; the new tables are independent empirical runs, not quantities forced by those citations or by definition of the training objective. Zero-fill missing-cue handling is an experimental protocol choice, not a circular reduction. The central claim—decoupled frontends plus heterogeneous batching enable multi-cue composition without redesign—is evaluated externally and does not collapse to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

Load-bearing content is engineering abstraction plus empirical SI-SDRi under chosen datasets and training recipes. Claims rest on standard deep TSE practice (cue-conditioned separation, SI-SDR training), simulated acoustics, and the modeling choice that missing cues can be zero-filled through fixed fusion interfaces. No formal derivation; free parameters are ordinary ML hyperparameters and architecture widths.

free parameters (4)
  • BSRNN subband/LSTM configuration = 32 subbands, dim 128, 6×192 BiLSTM
    32 subbands, 128-dim features, 192-dim BiLSTM stacked 6 times—hand-chosen backbone capacity that conditions all default comparisons.
  • Training schedule = 150 epochs, 3 s segments
    150 epochs, 3 s segments, Adam with a fixed but unspecified identical learning schedule; affects absolute SI-SDRi levels across tables.
  • Missing-cue rates in heterogeneous training = 30%/30%/~9%
    Enrollment absent 30%, spatial absent 30%, both ~9%—hand-set simulation of availability that drives Table 7 conclusions.
  • Room/SINR simulation ranges = RT60 U(0.1,0.5), SINR [−10,10] dB
    RT60 0.1–0.5 s, SINR −10 to 10 dB, overlap 0–100%, 4-mic ULA 10 cm—define the spatial-cue regime.
axioms (5)
  • domain assumption Target speech can be estimated as a function of the mixture plus a variable subset of encoded auxiliary cues (generalized TSE formulation Eqs. 1–5).
    Section 2.1; standard TSE premise extended to heterogeneous Cn.
  • ad hoc to paper Cue encoders and separator backbones can be fully decoupled via standardized injection interfaces without needing cue-specific redesign of f(·).
    Design principle in Sections 2.2 and 3.2–3.4; load-bearing for the modularity claim.
  • ad hoc to paper Missing cues may be represented as zero-valued placeholders while keeping the same architecture and loss.
    Section 4.6.2 heterogeneous training pipeline.
  • domain assumption Negative SI-SDR is an adequate training objective and SI-SDRi (with accuracy = fraction SI-SDRi>1 dB) adequately ranks cue configurations.
    Section 4.1; common in separation literature but not validated against perceptual or downstream ASR metrics here.
  • domain assumption Image-source simulated multi-channel mixtures sufficiently represent spatial cue conditions for comparing handcrafted vs embedding spatial features.
    Section 4.3 dataset construction via gpuRIR.
invented entities (1)
  • WeSep modular Top Model / heterogeneous sample abstraction no independent evidence
    purpose: Unify cue frontends, separator backbones, config-time fusion, and mixed-availability batching for TSE.
    Core software abstraction introduced to operationalize heterogeneous cue-conditioned TSE; prior WeSep lacked the generalized multi-cue interface.

pith-pipeline@v1.2.0-daily-grok45 · 13965 in / 3235 out tokens · 57563 ms · 2026-07-31T00:56:36.713848+00:00 · methodology

0 comments
read the original abstract

The study of Target Speaker Extraction (TSE) aims to isolate a desired speaker from overlapping speech mixture given auxiliary cues. Existing systems are typically designed for specific cue types, limiting flexibility when cue availability varies across scenarios. We present WeSep, a unified framework that reformulates TSE as a heterogeneous cue-conditioned learning problem. In WeSep, cue modules and separator backbones are decoupled through standardized interfaces, enabling configurable cue injection and flexible integration of diverse modalities. The design enables systematic study of cue structure, intra- and cross-modal interaction, and dynamic cue availability within a shared optimization framework, facilitating adaptation to real-world conditions. Experiments across enrollment, spatial, visual, and textual cues reveal modality-dependent characteristics and demonstrate stable optimization under heterogeneous cue availability. The toolkit will be publicly available.

Figures

Figures reproduced from arXiv: 2607.27436 by Haizhou Li, Haoyu Li, Ke Zhang, Shuai Wang, Shuhan Zhang, Xiaoyang Yu.

Figure 2
Figure 2. Figure 2: Modular Top Model construction in WeSep. A con￾figuration file instantiates the separator backbone and cue fron￾tends. Frontends may include multiple feature modules injected at different separator locations. Spatial cues are integrated into intermediate states, visual cues via standard fusion, and textual frontends may remain optional. The design enables decoupled cue modeling and flexible intra- and cros… view at source ↗
Figure 1
Figure 1. Figure 1: Heterogeneous data organization in WeSep. The samples.list defines mix–target pairs, and cue reposito￾ries are indexed by ID-spk or ID-mix-spk depending on modality. Solid links indicate consistently available cues (e.g., speaker, textual), while dashed links denote conditional cues (e.g., spatial, visual), reflecting heterogeneous availability. 2.2. Design Principles of Modular Cue-Composable TSE The gene… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 5 linked inside Pith

  1. [1]

    Deep learning has significantly improved separation performance un- der controlled conditions

    Introduction Speech separation (SS) aims to decompose a mixture into in- dividual sources and has been extensively studied [1, 2]. Deep learning has significantly improved separation performance un- der controlled conditions. However, conventional SS typically assumes a fixed number of outputs and lacks explicit target specification, making it difficult t...

  2. [2]

    Generalized Target Speaker Extraction Letx∈R T denote a mixture speech ands∈R T denote the tar- get speech in the mixture

    Problem Formulation and Design Motivation 2.1. Generalized Target Speaker Extraction Letx∈R T denote a mixture speech ands∈R T denote the tar- get speech in the mixture. In Target Speaker Extraction (TSE), the estimation ofsis conditioned on a set of auxiliary cues C={c 1, c2, . . . , cK },(1) which may correspond to enrollment speech, spatial informa- ti...

    Pith/arXiv arXiv 2026

  3. [3]

    It operationalizes the generalized TSE for- mulation in Section 2 and supports systematic exploration of single-, multi-, and heterogeneous-cue settings

    The WeSep Modular Framework WeSep is a modular and configurable framework that separates data abstraction, cue frontends, separator backbones, and top- level composition. It operationalizes the generalized TSE for- mulation in Section 2 and supports systematic exploration of single-, multi-, and heterogeneous-cue settings. Figure 1 illustrates the heterog...

  4. [4]

    Experiments 4.1. General Experimental Settings Unless otherwise specified, BSRNN [28] is adopted as the de- fault separator so that comparisons focus on cue configurations, including speaker, spatial, visual, textual, and their composi- tions. In the BSRNN separator, the frequency spectrum is di- vided into 32 subbands with a feature dimension of 128 per ...

  5. [5]

    Conclusion and Discussion This work revisits target speaker extraction as a problem of structured cue conditioning under dynamic availability. By ab- stracting cue representations from separator backbones, WeSep provides a unified experimental substrate for studying how cue granularity, modality interaction, and cue absence affect extrac- tion behavior. R...

  6. [6]

    62401377 and No

    Acknowledgments This research is supported by National Natural Science Foun- dation of China (Grant No. 62401377 and No. 62271432) and Yangtze River Delta Science and Technology Innovation Com- munity Joint Research Project (Grant No. 2024CSJGG1100) and Shenzhen Science and Technology Program (Shenzhen Key Laboratory Grant No. ZDSYS20230626091302006) and ...

    2023

  7. [7]

    Generative AI Use Disclosure Generative AI tools were used solely for language refinement during manuscript preparation and did not contribute to the core ideas, methodology, or experimental design of this work

  8. [8]

    Advances in speech separation: Techniques, challenges, and future trends,

    K. Li, G. Chen, W. Sang, Y . Luo, Z. Chen, S. Wang, S. He, Z.-Q. Wang, A. Li, Z. Wu, and X. Hu, “Advances in speech separation: Techniques, challenges, and future trends,” 2025. [Online]. Available: https://arxiv.org/abs/2508.10830

    Pith/arXiv arXiv 2025

  9. [9]

    Supervised speech separation based on deep learning: An overview,

    D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, vol. 26, no. 10, pp. 1702– 1726, 2018

    2018

  10. [10]

    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

    2023

  11. [11]

    Usef-tse: Universal speaker embedding free target speaker extraction,

    B. Zeng and M. Li, “Usef-tse: Universal speaker embedding free target speaker extraction,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 2110–2124, 2025

    2025

  12. [12]

    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,” inICASSP 2025 - 2025 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  13. [13]

    Listen to extract: Onset-prompted target speaker extraction,

    P. Shen, K. Chen, S. He, P. Chen, S. Yuan, H. Kong, X. Zhang, and Z.-Q. Wang, “Listen to extract: Onset-prompted target speaker extraction,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 4832–4843, 2025

    2025

  14. [14]

    3d spatial features for multi-channel target speech separation,

    R. Gu, S.-X. Zhang, M. Yu, and D. Yu, “3d spatial features for multi-channel target speech separation,” in2021 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 996–1002

    2021

  15. [15]

    3s-tse: Efficient three-stage target speaker extraction for real-time and low-resource applications,

    S. He, J. Liu, H. Li, Y . Yang, F. Chen, and X. Zhang, “3s-tse: Efficient three-stage target speaker extraction for real-time and low-resource applications,” inICASSP 2024 - 2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 421–425

    2024

  16. [16]

    Doa or speaker embed- ding: Which is better for multi-microphone target speaker extrac- tion,

    S. Zhang, J. Zhang, Y . Wang, and H. Yan, “Doa or speaker embed- ding: Which is better for multi-microphone target speaker extrac- tion,”IEEE Signal Processing Letters, vol. 32, pp. 3350–3354, 2025

    2025

  17. [17]

    Muse: Multi-modal tar- get speaker extraction with visual cues,

    Z. Pan, R. Tao, C. Xu, and H. Li, “Muse: Multi-modal tar- get speaker extraction with visual cues,” in2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6678–6682

    2021

  18. [18]

    Usev: Universal speaker extraction with visual cue,

    Z. Pan, M. Ge, and H. Li, “Usev: Universal speaker extraction with visual cue,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 3032–3045, 2022

    2022

  19. [19]

    Efficient audio-visual speech separa- tion with discrete lip semantics and multi-scale global-local atten- tion,

    K. Li, K. Gao, and X. Hu, “Efficient audio-visual speech separa- tion with discrete lip semantics and multi-scale global-local atten- tion,”arXiv preprint arXiv:2509.23610, 2025

    arXiv 2025

  20. [20]

    Detect, attend and extract: Keyword guided target speaker extraction,

    H. Li, Y . Xi, Y . Jiang, S. Wang, K. Knill, M. Gales, H. Li, and K. Yu, “Detect, attend and extract: Keyword guided target speaker extraction,”arXiv preprint arXiv:2602.07977, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    Contex- tual speech extraction: Leveraging textual history as an implicit cue for target speech extraction,

    M. Kim, R. Mira, H. Chen, S. Petridis, and M. Pantic, “Contex- tual speech extraction: Leveraging textual history as an implicit cue for target speech extraction,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2025, pp. 1–5

    2025

  22. [22]

    Speaker Extraction with Detection of Presence and Absence of Target Speakers,

    K. Zhang, M. Borsdorf, Z. Pan, H. Li, Y . Wei, and Y . Wang, “Speaker Extraction with Detection of Presence and Absence of Target Speakers,” inProc. Interspeech 2023, 2023, pp. 3714– 3718

    2023

  23. [23]

    Momuse: Momentum multi-modal target speaker extraction for real-time scenarios with impaired visual cues,

    J. Li, K. Zhang, S. Wang, K. Lee, M. Mak, and H. Li, “Momuse: Momentum multi-modal target speaker extraction for real-time scenarios with impaired visual cues,” in2025 IEEE International Conference on Multimedia and Expo, 2025

    2025

  24. [24]

    Multimodal attention fusion for target speaker ex- traction,

    H. Sato, T. Ochiai, K. Kinoshita, M. Delcroix, T. Nakatani, and S. Araki, “Multimodal attention fusion for target speaker ex- traction,” in2021 IEEE Spoken Language Technology Workshop (SLT), 2021, pp. 778–784

    2021

  25. [25]

    Beyond speaker identity: Text guided target speech ex- traction,

    M. Huo, A. Jain, C. P. Huynh, F. Kong, P. Wang, Z. Liu, and V . Bhat, “Beyond speaker identity: Text guided target speech ex- traction,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  26. [26]

    WeSep: A Scalable and Flexible Toolkit Towards Gen- eralizable 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 Gen- eralizable Target Speaker Extraction,” inInterspeech 2024, 2024, pp. 4273–4277

    2024

  27. [27]

    ClearerV oice-Studio: Bridging Ad- vanced Speech Processing Research and Practical Deployment,

    S. Zhao, Z. Pan, and B. Ma, “ClearerV oice-Studio: Bridging Ad- vanced Speech Processing Research and Practical Deployment,” inInterspeech 2025, 2025, pp. 2980–2984

    2025

  28. [28]

    Wespeaker: A research and production oriented speaker embedding 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 embedding learning toolkit,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  29. [29]

    ECAPA- TDNN: Emphasized Channel Attention, Propagation and Aggre- gation in TDNN Based Speaker Verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA- TDNN: Emphasized Channel Attention, Propagation and Aggre- gation in TDNN Based Speaker Verification,” inProc. Interspeech 2020, 2020, pp. 3830–3834

    2020

  30. [30]

    A study of multichannel spatiotemporal features and knowledge distillation on robust target speaker extraction,

    Y . Wang, J. Zhang, S. Chen, W. Zhang, Z. Ye, X. Zhou, and L. Dai, “A study of multichannel spatiotemporal features and knowledge distillation on robust target speaker extraction,” inICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 431–435

    2024

  31. [31]

    End-to-end doa-guided speech extraction in noisy multi-talker scenarios,

    K. Jing, W. Zhang, and Y . Gao, “End-to-end doa-guided speech extraction in noisy multi-talker scenarios,”arXiv preprint arXiv:2507.20926, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Spatially selective deep non-linear filters for speaker extraction,

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

    2023

  33. [33]

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

    J. Han, Y . Long, L. Burget, and J. ˇCernock`y, “Dpccn: Densely- connected pyramid complex convolutional network for robust speech separation and extraction,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2022, pp. 7292–7296

    2022

  34. [34]

    Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,

    Y . Luo and N. Mesgarani, “Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,”IEEE/ACM transactions on audio, speech, and language processing, vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  35. [35]

    Music source separation with band-split rnn,

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

    2023

  36. [36]

    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. Watan- abe, “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

  37. [37]

    Nbc2: Multichannel speech separation with revised narrow-band conformer,

    C. Quan and X. Li, “Nbc2: Multichannel speech separation with revised narrow-band conformer,” 2022. [Online]. Available: https://arxiv.org/abs/2212.02076

    Pith/arXiv arXiv 2022

  38. [38]

    Sdr– half-baked or well done?

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

    2019

  39. [39]

    Librimix: An open-source dataset for generalizable speech separation,

    J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vin- cent, “Librimix: An open-source dataset for generalizable speech separation,” 2020

    2020

  40. [40]

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

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  41. [41]

    100 nonspeech environmental sounds,

    G. Hu, “100 nonspeech environmental sounds,”The Ohio State University, Department of Computer Science and Engineering, 2004

    2004

  42. [42]

    gpurir: A python library for room impulse response simulation with gpu acceler- ation,

    D. Diaz-Guerra, A. Miguel, and J. R. Beltran, “gpurir: A python library for room impulse response simulation with gpu acceler- ation,”Multimedia Tools and Applications, vol. 80, no. 4, pp. 5653–5671, 2021

    2021