Pith. sign in

REVIEW 2 major objections 5 minor 35 references

LLM multi-talker ASR fails when acoustic evidence is only a one-shot prefix; persistent talker-structured memory fixes the three-talker collapse.

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

2026-07-13 17:08 UTC pith:DGCI55YI

load-bearing objection Solid systems paper: gated residual acoustic memory + two-stage LoRA closes most of the three-talker LLM-SOT gap with clean ablations; ordinary engineering soft spots only. the 2 major comments →

arxiv 2603.27205 v2 pith:DGCI55YI submitted 2026-03-28 cs.SD

Beyond Acoustic Prefixes: Persistent Grounding in Serialized Acoustic Memory for LLM-Based Multi-Talker Speech Recognition

Hao Shi , Yuan Gao , Xugang Lu , Tatsuya Kawahara This is my paper
classification cs.SD
keywords multi-talker ASRserialized output traininglarge language modelscross-attention adaptersserialized acoustic memoryLoRALibriMix
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.

Large language models work well as decoders for two-speaker overlapping speech recognition under serialized output training, but they break down once three speakers interleave. The paper argues the failure is not mainly a lack of linguistic skill; it is a conditioning bottleneck: conventional systems give the model acoustic evidence only as a static projected prefix, so the decoder must carry fine-grained who-spoke-when information through the entire generation process by itself. Richer prefixes built from CTC tokens or continuous acoustic cues help a little, yet still leave three-talker performance far below what is needed. The authors therefore keep talker-disentangled, onset-ordered acoustic representations as external memory and let every decoder layer re-query that memory through a gated residual cross-attention adapter, then refine both the acoustic pathway and the language-model self-attention with low-rank updates. On standard two- and three-talker mixtures the method yields consistent word-error reductions, with the largest gains precisely where ordinary LLM-SOT collapses.

Core claim

The central claim is that persistent, gated access to talker-disentangled and onset-ordered acoustic memory inside the LLM decoder, rather than one-shot prefix conditioning, is what restores reliable multi-talker recognition—especially under three-talker interleaving—and that a two-stage adaptation (gated residual cross-attention followed by joint LoRA refinement of the acoustic and self-attention pathways) makes this injection stable and effective for large backbones.

What carries the argument

Serialized acoustic memory accessed by gated residual cross-attention: talker-separated, onset-ordered encoder streams are retained as external key/value memory; each decoder layer attends to them and injects the result as a gated residual update so pretrained language representations are not overwritten.

Load-bearing premise

The method assumes the frozen separator and serialized CTC branches produce clean enough, correctly ordered talker streams that the gated adapters can usefully re-query throughout decoding; if those streams remain heavily mixed or mis-ordered, persistent grounding cannot recover the right assignment.

What would settle it

On a held-out three-talker test set, replace the talker-separated memory with deliberately mixed or randomly permuted streams while keeping the same gated adapters and LoRA refinement; if word error does not rise sharply relative to the properly ordered memory, the claim that persistent talker-structured grounding is the decisive factor is falsified.

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

If this is right

  • LLM-based multi-talker ASR should treat acoustic evidence as re-queryable external memory rather than a disposable prefix.
  • Three-talker mixtures become a practical operating regime for LLM decoders once talker-structured memory is available at every step.
  • Naive full cross-attention stacking is unnecessary and often harmful; a lightweight gated residual adapter plus low-rank refinement is sufficient and more stable.
  • The same two-stage recipe can be applied to other LLM speech tasks that currently suffer from lossy or static acoustic conditioning.

Where Pith is reading between the lines

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

  • If the bottleneck is really loss of fine-grained acoustic grounding rather than model capacity, further scaling of the language-model backbone alone will not close the three-talker gap without similar memory mechanisms.
  • The same gated-memory pattern may transfer to other multi-source generation settings (e.g., multi-instrument transcription or multi-speaker translation) where assignment and ordering must be maintained over long outputs.
  • A natural next stress test is four-or-more talker mixtures or highly asynchronous onsets, where residual leakage in the separator would be expected to dominate.

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

2 major / 5 minor

Summary. The paper studies why LLM-based Serialized Output Training (SOT) works well for two-talker ASR but degrades on three-talker mixtures. It first shows that CTC-derived prefix prompts with increasing acoustic content (token-only, hybrid, continuous separated streams) improve over plain SOT but remain insufficient on Libri3Mix. It then proposes persistent grounding: talker-disentangled, onset-ordered acoustic streams are kept as external memory and injected into every decoder layer via a gated residual cross-attention adapter (Stage 1), followed by joint LoRA refinement of the cross-attention pathway and selected LLM self-attention projections (Stage 2). Experiments on Libri2Mix/Libri3Mix (clean/noisy) across LLaMA 1B/3B/8B base and instruct backbones report consistent WER reductions over LLM-SOT and naïve stacked cross-attention, with the largest gains in three-talker conditions.

Significance. The work addresses a concrete, practically important failure mode of LLM-SOT multi-talker ASR. The staged design is carefully controlled: prefix variants, stacked vs. gated adapters, and LoRA rank/alpha choices are ablated across three model scales and clean/noisy conditions (Tables I–IV), and the best systems are competitive with or better than strong non-LLM and serialized-CTC references on Libri3Mix (Table V). The gated residual formulation and post-merge inference design are simple and transferable. Strengths include extensive empirical coverage, clear isolation of the conditioning bottleneck, and parameter-efficient adaptation that does not add inference branches after merging. The central claim is empirical rather than theoretical, but the evidence is proportionate and reproducible in principle from the described pipeline.

major comments (2)
  1. The training pipeline in §V-C freezes the WavLM encoder after the initial SOT stage and then freezes the separator/SerCTC modules before Stage 1. Tables I–III show that serialized CTC alone is already a strong three-talker reference (e.g., Exp. 6/15/22/27/32/37). The paper should quantify how much of the final gain is attributable to better use of those frozen streams versus residual capacity left in the frozen encoder/separator. A short ablation that unfreezes the separator (or jointly trains SerCTC with Stage 1) would clarify whether the gated adapter is the main lever or whether the frozen EncSep front-end is the dominant factor.
  2. §VI-E and §VI-F note that a single fixed adapter configuration (Da=512, shared gate init) becomes less effective as backbone size grows to 8B, and that Stage-2 refinement still trails serialized CTC under noisy 8B conditions. Given that the abstract claims “consistent improvements … with particularly large gains in three-talker mixtures,” the manuscript should either (i) report backbone-specific Da / learning-rate scaling for 8B or (ii) explicitly qualify that the largest absolute gains and best absolute WER occur at 3B under the fixed recipe (Tables II, V). Without that qualification the scaling claim is slightly overstated.
minor comments (5)
  1. Title/abstract use “Beyond Acoustic Prefixes: Persistent Grounding…”, while the manuscript header and body title are “Two-Stage Acoustic Adaptation with Gated Cross-Attention Adapters…”. Align the title throughout.
  2. Eqs. (25)–(26) and Fig. 1 refer to E_aco / Projm as external memory, but several table rows still list prefix prompt as [Hp] even when cross-attention is used. Clarify in the table caption whether Hp remains the LLM prefix while E_aco is only the cross-attention memory.
  3. Typographical issues: “inforamtion” (Fig. 2 caption), “Comparision” (§VI-G), and inconsistent hyphenation of “cross-attention” / “cross attention”.
  4. §V-A states that three-talker offset files “will be released after the anonymous review period.” For reproducibility, note the release plan or provide a short description of the onset-time sampling procedure.
  5. Table V mixes prior SOP/SOT numbers from [15] with the present IDs; a one-sentence note that those earlier numbers used a different training recipe would avoid direct numerical over-interpretation.

Circularity Check

0 steps flagged

No significant circularity: empirical engineering paper whose WER gains are measured on held-out LibriMix, not forced by definition or self-cited uniqueness.

full rationale

The paper's load-bearing claims are comparative WER results (Tables I–V) for gated residual cross-attention plus Stage-2 LoRA refinement versus LLM-SOT, richer prefixes, and naïve stacked cross-attention. These are ordinary supervised evaluations on Libri2Mix/Libri3Mix test sets under clean and noisy conditions; nothing is derived as a first-principles prediction that reduces to a fitted constant or to an equation that is true by construction. Self-citations to EncSep, SOP, and serialized CTC supply the frozen separator/memory construction and strong non-LLM references, which is normal incremental engineering and is not used as a uniqueness theorem that forbids alternatives. The central architectural claim (persistent access to talker-ordered acoustic memory helps three-talker decoding) is falsifiable by the reported ablations and does not collapse into its inputs. Score 1 only for the minor, non-load-bearing self-citation of the separator pipeline; no circular step of kinds 1–6 is present.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

Empirical ML paper; load-bearing elements are architectural choices and training-stage freezes rather than mathematical axioms. Free parameters are the usual adapter hyper-parameters; invented entities are the specific gated residual adapter and the serialized acoustic memory construction.

free parameters (4)
  • cross-attention dimension Da = 512
    Fixed at 512 for all backbone sizes; authors note it is likely sub-optimal for 8B yet not retuned.
  • gate logit initialization γ = σ(−2)≈0.12
    Set so σ(γ)≈0.12 to start near the pretrained decoder; controls initial acoustic injection strength.
  • LoRA rank r and alpha = r=8, α=4 (best reported)
    Selected via small grid (Table IV); final reported systems use r=8, α=4 for joint refinement.
  • SerCTC trade-off α and separator LSTM size = hidden=796
    Inherited from prior EncSep work; separator is two-layer LSTM hidden 796.
axioms (3)
  • domain assumption Talker streams ordered by utterance onset time are a valid serialization for both CTC targets and acoustic memory.
    Standard SOT assumption (Section II-B); if onset order is ambiguous the memory keys become misaligned.
  • ad hoc to paper Freezing the WavLM encoder after the initial SOT stage preserves representations useful for both SerCTC and later cross-attention.
    Explicit training recipe in Section V-C; untested whether joint end-to-end fine-tuning would be superior.
  • ad hoc to paper Gated residual update H_out = H + g·Δ preserves pretrained language modeling capacity better than full replacement.
    Core design claim of Stage 1 (Eq. 31); supported by ablation vs stacked CA but not proven in general.
invented entities (2)
  • Gated residual cross-attention adapter no independent evidence
    purpose: Inject talker-aware acoustic memory into every LLM layer without overwriting pretrained self-attention states.
    Defined in Section IV-C (Eqs. 28–31); the scalar gate and residual form are the paper’s main architectural novelty.
  • Serialized acoustic memory E_aco no independent evidence
    purpose: Provide persistent, onset-ordered, talker-disentangled key/value store for the adapters.
    Constructed by concatenating separator outputs and projecting (Eqs. 25–26); distinct from ordinary mixture prefix Hp.

pith-pipeline@v1.1.0-grok45 · 24774 in / 2911 out tokens · 31871 ms · 2026-07-13T17:08:08.019729+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) are effective decoders for Serialized Output Training (SOT) in two-talker automatic speech recognition (ASR), but their performance degrades substantially in three-talker mixtures. A key limitation is that conventional systems provide acoustic evidence only through an initial projected prefix, requiring the decoder to preserve fine-grained talker information throughout autoregressive generation. We first revisit CTC-derived static prefix conditioning using discrete token, hybrid token-acoustic, and continuous acoustic prompts. Although continuous acoustic cues are more reliable than discrete CTC hypotheses, the improvements on Libri3Mix remain limited, showing that richer prefix content alone does not resolve the conditioning bottleneck. We therefore propose persistent grounding in serialized acoustic memory, which enables the decoder to retrieve talker-structured acoustic evidence throughout the utterance. Specifically, talker-disentangled and onset-ordered acoustic representations are retained as external memory and accessed during decoding through gated residual cross-attention. We further introduce joint low-rank refinement of the acoustic retrieval pathway and selected LLM self-attention projections using LoRA. Experiments on Libri2Mix and Libri3Mix under clean and noisy conditions show consistent improvements over conventional LLM-SOT and naive stacked cross-attention, with particularly large gains in three-talker mixtures. These results demonstrate the importance of persistent access to structured, talker-aware acoustic evidence for LLM-based multi-talker ASR.

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

35 extracted references · 5 linked inside Pith

  1. [1]

    End-to-end multi-talker overlapping speech recognition,

    A. Tripathi, H. Lu, and H. Sak, “End-to-end multi-talker overlapping speech recognition,” inProc. ICASSP, 2020, pp. 6129–6133

    2020

  2. [2]

    The chime-8 dasr challenge for generalizable and array agnostic distant automatic speech recognition and diarization,

    S. Cornell, T. J. Park, H. Huang, C. Boeddeker, X. Chang, M. Ma- ciejewski, M. S. Wiesner, P. Garcia, and S. Watanabe, “The chime-8 dasr challenge for generalizable and array agnostic distant automatic speech recognition and diarization,” in8th International Workshop on Speech Processing in Everyday Environments (CHiME 2024), 2024, pp. 1–6

    2024

  3. [3]

    Recent trends in distant conversational speech recognition: A review of CHiME-7 and 8 DASR challenges,

    S. Cornell, C. Boeddeker, T. Park, H. Huang, D. Raj, M. Wiesner, Y . Ma- suyama, X. Chang, Z.-Q. Wang, S. Squartini, P. Garcia, and S. Watanabe, “Recent trends in distant conversational speech recognition: A review of CHiME-7 and 8 DASR challenges,”arXiv preprint arXiv:2507.18161, 2025

    arXiv 2025

  4. [4]

    Waveform-domain speech enhancement using spectrogram encoding for robust speech recognition,

    H. Shi, M. Mimura, and T. Kawahara, “Waveform-domain speech enhancement using spectrogram encoding for robust speech recognition,” IEEE/ACM TASLP, vol. 32, pp. 3049–3060, 2024

    2024

  5. [5]

    An overview of noise- robust automatic speech recognition,

    J. Li, L. Deng, Y . Gong, and R. Haeb-Umbach, “An overview of noise- robust automatic speech recognition,”IEEE/ACM TASLP, vol. 22, no. 4, pp. 745–777, 2014

    2014

  6. [6]

    Machine learning paradigms for speech recognition: An overview,

    L. Deng and X. Li, “Machine learning paradigms for speech recognition: An overview,”IEEE TASLP, vol. 21, no. 5, pp. 1060–1089, 2013

    2013

  7. [7]

    Language-specific Characteristic Assistance for Code-switching Speech Recognition,

    T. Song, Q. Xu, M. Ge, L. Wang, H. Shi, Y . Lv, Y . Lin, and J. Dang, “Language-specific Characteristic Assistance for Code-switching Speech Recognition,” inProc. Interspeech, 2022, pp. 3924–3928

    2022

  8. [8]

    End-to-end monaural multi-speaker ASR system without pretraining,

    X. Chang, Y . Qian, K. Yu, and S. Watanabe, “End-to-end monaural multi-speaker ASR system without pretraining,” inProc. ICASSP, 2019, pp. 6256–6260

    2019

  9. [9]

    End- to-end multi-speaker speech recognition,

    S. Settle, J. L. Roux, T. Hori, S. Watanabe, and J. R. Hershey, “End- to-end multi-speaker speech recognition,” inProc. ICASSP, 2018, pp. 4819–4823

    2018

  10. [10]

    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 TASLP, vol. 26, no. 10, pp. 1702– 1726, 2018

    2018

  11. [11]

    Investigation of speech separation as a front-end for noise robust speech recognition,

    A. Narayanan and D. Wang, “Investigation of speech separation as a front-end for noise robust speech recognition,”IEEE/ACM TASLP, vol. 22, no. 4, pp. 826–835, 2014

    2014

  12. [12]

    Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks,

    M. Kolbæk, D. Yu, Z.-H. Tan, and J. Jensen, “Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks,”IEEE/ACM TASLP, vol. 25, no. 10, pp. 1901–1913, 2017

    1901

  13. [13]

    Serialized Output Training for End-to-End Overlapped Speech Recognition,

    N. Kanda, Y . Gaur, X. Wang, Z. Meng, and T. Yoshioka, “Serialized Output Training for End-to-End Overlapped Speech Recognition,” in Proc. Interspeech, 2020, pp. 2797–2801

    2020

  14. [14]

    Ba- sot: Boundary-aware serialized output training for multi-talker asr,

    Y . Liang, F. Yu, Y . Li, P. Guo, S. Zhang, Q. Chen, and L. Xie, “Ba- sot: Boundary-aware serialized output training for multi-talker asr,” in INTERSPEECH, 2023, pp. 3487–3491

    2023

  15. [15]

    Serialized output prompting for large language model-based multi- talker speech recognition,

    H. Shi, Y . Fujita, T. Mizumoto, L. Liu, A. Kojima, and Y . Sudo, “Serialized output prompting for large language model-based multi- talker speech recognition,”arXiv preprint arXiv:2509.04488, 2025. [Online]. Available: https://arxiv.org/abs/2509.04488

    Pith/arXiv arXiv 2025

  16. [16]

    Advancing multi-talker ASR performance with large lan- guage models,

    M. Shi, Z. Jin, Y . Xu, Y . Xu, S.-X. Zhang, K. Wei, Y . Shao, C. Zhang, and D. Yu, “Advancing multi-talker ASR performance with large lan- guage models,” inProc. SLT, 2024, pp. 14–21

    2024

  17. [17]

    Large language model can transcribe speech in multi-talker scenarios with versatile instructions,

    L. Meng, S. Hu, J. Kang, Z. Li, Y . Wang, W. Wu, X. Wu, X. Liu, and H. Meng, “Large language model can transcribe speech in multi-talker scenarios with versatile instructions,” inProc. ICASSP, 2025, pp. 1–5

    2025

  18. [18]

    Serialized speech information guidance with overlapped encoding separation for multi-speaker auto- matic speech recognition,

    H. Shi, Y . Gao, Z. Ni, and T. Kawahara, “Serialized speech information guidance with overlapped encoding separation for multi-speaker auto- matic speech recognition,” inProc. SLT, 2024, pp. 198–204

    2024

  19. [19]

    Wavlm: Large-scale self-supervised pre- training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “Wavlm: Large-scale self-supervised pre- training for full stack speech processing,”IEEE JSTSP, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  20. [20]

    Cocktail hubert: Generalized self- supervised pre-training for mixture and single-source speech,

    M. Fazel-Zarandi and W.-N. Hsu, “Cocktail hubert: Generalized self- supervised pre-training for mixture and single-source speech,” inProc. ICASSP, 2023, pp. 1–5

    2023

  21. [21]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  22. [22]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  23. [23]

    Exploration of adapter for noise robust automatic speech recognition,

    H. Shi and T. Kawahara, “Exploration of adapter for noise robust automatic speech recognition,”arXiv preprint arXiv:2402.18275, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Dual-path adaptation of pretrained feature extraction module for robust automatic speech recognition,

    ——, “Dual-path adaptation of pretrained feature extraction module for robust automatic speech recognition,” inProc. Interspeech, 2024, pp. 2850–2854

    2024

  25. [25]

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

    J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vincent, “Librimix: An open-source dataset for generalizable speech separation,” arXiv preprint arXiv:2005.11262, 2020

    Pith/arXiv arXiv 2005

  26. [26]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public domain audio books,” inProc. ICASSP, 2015, pp. 5206–5210

    2015

  27. [27]

    Wham!: Extending speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. Le Roux, “Wham!: Extending speech separation to noisy environments,” inProc. Interspeech, Sep. 2019

    2019

  28. [28]

    Distilling llm semantic priors into encoder-only multi-talker asr with talker-count routing,

    H. Shi, Y . Fujita, R. Koshkin, M. Zhao, Y . Gao, L. Liu, and Y . Sudo, “Distilling llm semantic priors into encoder-only multi-talker asr with talker-count routing,”arXiv preprint arXiv:2603.10587, 2026

    arXiv 2026

  29. [29]

    Exploring time-frequency domain target speaker extraction for causal and non-causal processing,

    W. Zhang, L. Yang, and Y . Qian, “Exploring time-frequency domain target speaker extraction for causal and non-causal processing,” inProc. ASRU, 2023, pp. 1–6

    2023

  30. [30]

    Unifying diarization, separation, and asr with multi-speaker encoder,

    M. Shakeel, Y . Sudo, Y . Peng, C.-J. Lin, and S. Watanabe, “Unifying diarization, separation, and asr with multi-speaker encoder,”arXiv preprint arXiv:2508.20474, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Hypothesis clustering and merging: Novel multitalker speech recog- nition with speaker tokens,

    Y . Kashiwagi, H. Futami, E. Tsunoo, S. Arora, and S. Watanabe, “Hypothesis clustering and merging: Novel multitalker speech recog- nition with speaker tokens,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  32. [32]

    Solospeech: Enhancing intelligibility and quality in target speech extraction through a cascaded generative pipeline,

    H. Wang, J. Hai, D. Yang, C. Chen, K. Li, J. Peng, T. Thebaud, L. M. Velazquez, J. Villalba, and N. Dehak, “Solospeech: Enhancing intelligibility and quality in target speech extraction through a cascaded generative pipeline,”arXiv preprint arXiv:2505.19314, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    A sidecar separator can convert a single-talker speech recognition system to a multi-talker one,

    L. Meng, J. Kang, M. Cui, Y . Wang, X. Wu, and H. Meng, “A sidecar separator can convert a single-talker speech recognition system to a multi-talker one,” inProc. ICASSP, 2023, pp. 1–5

    2023

  34. [34]

    Adapting self- supervised models to multi-talker speech recognition using speaker embeddings,

    Z. Huang, D. Raj, P. Garc ´ıa, and S. Khudanpur, “Adapting self- supervised models to multi-talker speech recognition using speaker embeddings,” inProc. ICASSP, 2023, pp. 1–5

    2023

  35. [35]

    Enhancing target-speaker automatic speech recognition using multiple speaker embedding extractors with virtual speaker embedding,

    J.-S. Seong, J.-H. Choi, Y .-R. Jeoung, I. Kim, and J.-H. Chang, “Enhancing target-speaker automatic speech recognition using multiple speaker embedding extractors with virtual speaker embedding,” inProc. INTERSPEECH, 2025, pp. 4918–4922

    2025