pith. sign in

arxiv: 2606.13095 · v1 · pith:3LLCW3RUnew · submitted 2026-06-11 · 📡 eess.AS · cs.SD

Balancing ASR and diarization in end-to-end LLMs for multi-talker speech recognition

Naijun Zheng , Yuke Lin , Sanli Tian , Mengtian Li , Zhiwei Lin , Longshuai Xiao , Dandan Tu This is my paper

Pith reviewed 2026-06-27 06:02 UTC · model grok-4.3

classification 📡 eess.AS cs.SD
keywords multi-talker speech recognitionspeaker diarizationend-to-end LLMASRlimited training datadual-encoder architecturefeature interleaving
0
0 comments X

The pith

Four strategies let an end-to-end LLM handle multi-talker speech recognition after training on limited real-recorded data.

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

This paper investigates how to train an LLM-based system for multi-talker speech recognition without relying on large-scale annotated corpora. It introduces a dual-encoder architecture to extract separate semantic and speaker features, a feature interleaving format to feed them into the LLM, a length-aware speaker ID loss, and an adaptive threshold on the ASR loss. These changes are intended to keep the model from favoring one task over the other during training on scarce real data. A sympathetic reader would care because expensive multi-talker annotation has been the main practical barrier to joint semantic-and-speaker modeling.

Core claim

The central claim is that the four proposed strategies—a dual-encoder architecture, feature interleaving format, length-aware speaker ID loss, and adaptive threshold for ASR loss—balance ASR and diarization training in an LLM-based multi-talker system. This balance allows high speaker-attribution accuracy when only limited real-recorded data are available, producing relative improvements of 18 percent on the AliMeeting corpus and 24 percent on the Aishell4 corpus over open-source baselines.

What carries the argument

The dual-encoder architecture that extracts semantic and speaker features separately, combined with their interleaving as LLM input and the two task-specific loss adjustments.

If this is right

  • The dual-encoder separates semantic content from speaker identity before the LLM sees the input.
  • Feature interleaving lets the single LLM attend to both kinds of information in one forward pass.
  • The length-aware speaker ID loss strengthens diarization without harming word recognition.
  • The adaptive threshold on ASR loss reduces hallucinations that arise from overlapping speech.
  • The resulting system exceeds open-source pipeline and LLM baselines on two standard multi-talker corpora.

Where Pith is reading between the lines

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

  • The same loss-balancing tactics might reduce dependence on synthetic multi-talker data in other joint audio tasks.
  • If the interleaving format works for two feature streams, it could be tested on additional modalities such as visual or textual context.
  • Real-world deployments with only small collections of meeting recordings could become feasible sooner than data-hungry alternatives.

Load-bearing premise

The four strategies together are sufficient to keep ASR and diarization training in balance on limited real-recorded data while preserving speaker attribution accuracy.

What would settle it

Reproducing the system on the AliMeeting or Aishell4 test sets and finding no relative improvement, or a drop, in combined ASR and diarization metrics would show the strategies do not achieve the claimed balance.

Figures

Figures reproduced from arXiv: 2606.13095 by Dandan Tu, Longshuai Xiao, Mengtian Li, Naijun Zheng, Sanli Tian, Yuke Lin, Zhiwei Lin.

Figure 1
Figure 1. Figure 1: Dual-encoder architecture with different feature fu￾sion strategies. the input sequence length. To balance robustness and tempo￾ral resolution, we use three chunk sizes (400 ms, 200 ms, and 100 ms). Longer chunks yield more stable speaker embeddings, while shorter chunks provide finer time granularity. The embed￾dings from all chunk sizes are concatenated along the feature axis and passed through projectio… view at source ↗
Figure 2
Figure 2. Figure 2: CE loss within overlaps, where the tokens in overlapped regions are of high ASR loss. (Segmented from R003 M0046 in the Alimeeting training set) We investigate the causes of hallucinations during train￾ing [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: System performance with different downsampling rates applied to speaker feature sequences. pleteness. By contrast, temporal merging approaches preserve speaker information more effectively. In particular, the tempo￾ral interleave format, which relies on positional encoding be￾tween adjacent segments, aligns more naturally with our label structure and achieves the best performance on ∆cp. Applying the adapt… view at source ↗
read the original abstract

Multi-talker speech recognition is often addressed by combining automatic speech recognition (ASR) and speaker diarization in a pipeline system. Recently, LLM-based approaches have shown promise by jointly modeling semantic and speaker information, but they typically require large-scale multi-talker corpora that are costly to annotate. In this paper, we investigate how to efficiently train an LLM-based system with limited real-recorded data while maintaining high accuracy in speaker attribution. We propose several strategies: (1) a dual-encoder architecture to extract semantic and speaker features, (2) a feature interleaving format to merge these features as the inputs to the LLM, (3) a length-aware speaker ID loss to enhance diarization capability, and (4) an adaptive threshold strategy for ASR loss computation to mitigate hallucinations caused by speech overlaps. These strategies balance training between ASR and diarization tasks. Our system outperforms open-source baseline approaches, achieving relative improvements of 18% on the AliMeeting corpus and 24% on the Aishell4 corpus.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes four strategies—a dual-encoder architecture, feature interleaving format, length-aware speaker ID loss, and adaptive threshold for ASR loss—to train an end-to-end LLM-based multi-talker ASR system on limited real-recorded data. It claims these balance ASR and diarization objectives while mitigating overlap-induced hallucinations, yielding relative WER improvements of 18% on AliMeeting and 24% on Aishell4 over open-source baselines.

Significance. If the empirical gains are shown to be robust and attributable to the proposed components, the work would be significant for practical multi-talker ASR, as it addresses the data-efficiency bottleneck that currently limits LLM-based joint modeling approaches. The explicit focus on balancing the two tasks via a heuristic threshold is a concrete contribution, though its generality remains to be verified.

major comments (3)
  1. [Experiments] Experiments section (results tables): No ablation is reported that removes only the adaptive threshold while keeping the dual-encoder, interleaving, and length-aware loss fixed. Without overlap-specific metrics (e.g., WER on overlap vs. non-overlap subsets) or this controlled ablation, it is impossible to confirm that the threshold drives the claimed mitigation of hallucinations or the 18%/24% gains, as opposed to the structural changes alone. This directly undermines the central claim that the four strategies together suffice.
  2. [Results] Results section: Relative improvements are stated without error bars, standard deviations across runs, or statistical significance tests against the baselines. Given that the abstract presents these numbers as the primary evidence, the absence of variability measures makes it impossible to judge whether the reported gains are reliable or could arise from training stochasticity.
  3. [§3.4] §3.4 (adaptive threshold description): The threshold is described as mitigating hallucinations from overlaps, yet the manuscript provides no quantitative analysis of how the threshold value is chosen or its sensitivity. If the threshold is a key load-bearing heuristic, its derivation or validation procedure should be shown explicitly rather than left as an empirical choice.
minor comments (2)
  1. [Abstract] The abstract and introduction refer to 'open-source baseline approaches' without naming the specific systems or their configurations in the main text; this should be clarified for reproducibility.
  2. [§3] Notation for the feature interleaving format and the length-aware speaker ID loss could be made more explicit (e.g., by defining the input tensor shapes) to aid readers implementing the dual-encoder.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments that help clarify the contributions of our proposed strategies. We address each major point below and will revise the manuscript accordingly to strengthen the empirical support for our claims.

read point-by-point responses

  1. Referee: [Experiments] Experiments section (results tables): No ablation is reported that removes only the adaptive threshold while keeping the dual-encoder, interleaving, and length-aware loss fixed. Without overlap-specific metrics (e.g., WER on overlap vs. non-overlap subsets) or this controlled ablation, it is impossible to confirm that the threshold drives the claimed mitigation of hallucinations or the 18%/24% gains, as opposed to the structural changes alone. This directly undermines the central claim that the four strategies together suffice.

    Authors: We agree that a controlled ablation isolating the adaptive threshold is essential to attribute gains specifically to it rather than the architectural changes. In the revised version, we will add this ablation (training without the threshold while fixing the other three components) and report WER on overlap versus non-overlap subsets to directly demonstrate the threshold's role in reducing hallucinations. revision: yes

  2. Referee: [Results] Results section: Relative improvements are stated without error bars, standard deviations across runs, or statistical significance tests against the baselines. Given that the abstract presents these numbers as the primary evidence, the absence of variability measures makes it impossible to judge whether the reported gains are reliable or could arise from training stochasticity.

    Authors: We acknowledge that variability measures are needed to assess reliability. Due to the high computational cost of LLM training, only single runs were performed. In revision we will add multiple runs with different seeds where feasible and report standard deviations; otherwise we will explicitly note this limitation while highlighting consistency across the two evaluation corpora. revision: partial

  3. Referee: [§3.4] §3.4 (adaptive threshold description): The threshold is described as mitigating hallucinations from overlaps, yet the manuscript provides no quantitative analysis of how the threshold value is chosen or its sensitivity. If the threshold is a key load-bearing heuristic, its derivation or validation procedure should be shown explicitly rather than left as an empirical choice.

    Authors: We will expand §3.4 with a quantitative description of the threshold selection process, including the validation procedure on a held-out set and a sensitivity analysis showing WER variation across different threshold values. This will make the empirical choice explicit and demonstrate its robustness. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical method proposal with external baselines

full rationale

The paper proposes four training strategies (dual-encoder, feature interleaving, length-aware speaker ID loss, adaptive threshold) and reports relative WER improvements of 18% and 24% versus open-source baselines on AliMeeting and Aishell4. No derivation chain, first-principles result, fitted parameter renamed as prediction, or self-citation load-bearing step exists. Performance numbers are direct empirical comparisons to external systems; the strategies are architectural/heuristic choices whose efficacy is tested by ablation or comparison outside the paper's own fitted values. The central claim therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical model is present; the paper is an empirical ML engineering study whose claims rest on architectural choices and loss modifications rather than axioms or fitted constants.

pith-pipeline@v0.9.1-grok · 5733 in / 1135 out tokens · 26191 ms · 2026-06-27T06:02:57.032532+00:00 · methodology

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

37 extracted references · 1 linked inside Pith

  1. [1]

    These sys- tems are relatively lightweight in parameters and deliver robust performance

    Introduction Multi-talker speech recognition[1, 2, 3] has been widely applied in scenarios such as meeting recordings, subtitle transcription, and dialogue understanding, aiming to solve the problem of ”who spoke what.” Since the task can be split into both au- tomatic speech recognition (ASR) and speaker identification tasks, it can be solved with the pi...

  2. [2]

    en”, “yes

    Proposed Method 2.1. Dual-encoder structure and feature fusion strategies Figure 1 illustrates the overall architecture, which consists of two encoders: an ASR encoder and a speaker-feature encoder. Given an audio segment containing multiple speakers, the dual encoders extract semantic features and speaker features, which are then combined in various form...

    Pith/arXiv arXiv 2026

  3. [3]

    Experiments In the first training stage, we use the WenetSpeech corpus [21]. For the second and third stages, we adopt an internal ASR corpus of approximately 4,000 hours, which consists of long- duration two-speaker conversations with transcriptions but does not contain overlap labels. Long recordings are truncated into shorter segments (8–20 seconds) ac...

  4. [4]

    The pipeline systems, whether using 3D-Speaker or DiariZen-large for diarization, achieve compa- rable results

    Results Table 2 presents the performance of different systems across three evaluation sets. The pipeline systems, whether using 3D-Speaker or DiariZen-large for diarization, achieve compa- rable results. In contrast, the end-to-end model VibeV oice-ASR struggles on highly overlapped segments. Our systems, built upon the SenseV oice-small encoder, in- heri...

    arXiv

  5. [5]

    The relationship between ASR and diarization tasks is carefully balanced through both the model architecture and the training scheme

    Conclusions In this paper, we proposed an end-to-end model for multi- speaker ASR with only 0.7B parameters, which achieves strong performance on far-field recordings with high overlap ratios. The relationship between ASR and diarization tasks is carefully balanced through both the model architecture and the training scheme. Our experiments demonstrate th...

  6. [6]

    These tools were not employed to gen- erate scientific ideas, experimental data, or technical contribu- tions

    Generative AI Use Disclosure During the preparation of this manuscript, generative AI tools were used solely for language editing and manuscript polishing to enhance readability. These tools were not employed to gen- erate scientific ideas, experimental data, or technical contribu- tions. All authors have thoroughly reviewed and approved the final version...

  7. [7]

    Deep neu- ral networks for single-channel multi-talker speech recognition,

    C. Weng, D. Yu, M. L. Seltzer, and J. Droppo, “Deep neu- ral networks for single-channel multi-talker speech recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 23, no. 10, pp. 1670–1679, 2015

    2015

  8. [8]

    Improv- ing end-to-end single-channel multi-talker speech recognition,

    W. Zhang, X. Chang, Y . Qian, and S. Watanabe, “Improv- ing end-to-end single-channel multi-talker speech recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 28, pp. 1385–1394, 2020

    2020

  9. [9]

    Integration of speech separation, diariza- tion, and recognition for multi-speaker meetings: System descrip- tion, comparison, and analysis,

    D. Raj, P. Denisov, Z. Chen, H. Erdogan, Z. Huang, M. He, S. Watanabe, J. Du, T. Yoshioka, Y . Luo, N. Kanda, J. Li, S. Wis- dom, and J. R. Hershey, “Integration of speech separation, diariza- tion, and recognition for multi-speaker meetings: System descrip- tion, comparison, and analysis,” in2021 IEEE Spoken Language Technology Workshop (SLT), 2021, pp. 897–904

    2021

  10. [10]

    Transcribe-to-diarize: Neural speaker diarization for unlimited number of speakers using end- to-end speaker-attributed asr,

    N. Kanda, X. Xiao, Y . Gaur, X. Wang, Z. Meng, Z. Chen, and T. Yoshioka, “Transcribe-to-diarize: Neural speaker diarization for unlimited number of speakers using end- to-end speaker-attributed asr,”ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8082–8086, 2021. [Online]. Available: https://api.se...

    2022

  11. [11]

    SCDiar: a streaming diarization system based on speaker change detection and speech recognition,

    N. Zheng, X. Wan, K. Liu, and Z. Huan, “SCDiar: a streaming diarization system based on speaker change detection and speech recognition,” inICASSP 2025 - 2025 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  12. [12]

    META-CAT: Speaker-informed speech embeddings via meta information con- catenation for multi-talker asr,

    J. Wang, W. Wang, K. Dhawan, T. Park, M. Kim, I. Medennikov, H. Huang, N. Koluguri, J. Balam, and B. Ginsburg, “META-CAT: Speaker-informed speech embeddings via meta information con- catenation for multi-talker asr,” inICASSP 2025 - 2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2025, pp. 1–5

    2025

  13. [13]

    Separate-to- recognize: Joint multi-target speech separation and speech recog- nition for speaker-attributed asr,

    Y . Lin, Z. Du, S. Zhang, F. Yu, Z. Zhao, and F. Wu, “Separate-to- recognize: Joint multi-target speech separation and speech recog- nition for speaker-attributed asr,” in2022 13th International Sym- posium on Chinese Spoken Language Processing (ISCSLP), 2022, pp. 150–154

    2022

  14. [14]

    DiarizationLM: Speaker Diarization Post-Processing with Large Language Models,

    Q. Wang, Y . Huang, G. Zhao, E. Clark, W. Xia, and H. Liao, “DiarizationLM: Speaker Diarization Post-Processing with Large Language Models,” inInterspeech 2024, 2024, pp. 3754–3758

    2024

  15. [15]

    Sortformer: A novel approach for permutation-resolved speaker supervision in speech-to-text systems,

    T. Park, I. Medennikov, K. Dhawan, W. Wang, H. Huang, N. R. Koluguri, K. C. Puvvada, J. Balam, and B. Ginsburg, “Sortformer: A novel approach for permutation-resolved speaker supervision in speech-to-text systems,” inProceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, A. Singh, M. Fazel, D. ...

  16. [16]

    48 153–48 169

    PMLR, 13–19 Jul 2025, pp. 48 153–48 169

    2025

  17. [17]

    Diarization-aware multi-speaker automatic speech recognition via large language models,

    Y . Lin, M. Cheng, Z. Li, B. Tang, and M. Li, “Diarization-aware multi-speaker automatic speech recognition via large language models,” 2025

    2025

  18. [18]

    Train short, infer long: Speech-llm enables zero-shot streamable joint asr and diarization on long audio,

    M. Shi, X. Xiao, R. Fan, S. Ling, and J. Li, “Train short, infer long: Speech-llm enables zero-shot streamable joint asr and diarization on long audio,” 2025. [Online]. Available: https://arxiv.org/abs/2511.16046

    arXiv 2025

  19. [19]

    TagSpeech: End-to-end multi- speaker asr and diarization with fine-grained temporal grounding,

    M. Huo, Y . Shao, and Y . Zhang, “TagSpeech: End-to-end multi- speaker asr and diarization with fine-grained temporal grounding,”

  20. [20]

    Available: https://arxiv.org/abs/2601.06896

    [Online]. Available: https://arxiv.org/abs/2601.06896

    arXiv

  21. [21]

    MOSS transcribe diarize technical report,

    M. AI, D. Yu, Z. Lin, C. Yang, Y . Zhang, H. Chen, J. Chen, K. Chen, L. Fan, Y . Jiang, J. Zhu, M. Li, W. Wang, Y . Wang, Z. Xu, Y . Gong, Y . Zhang, W. Zhang, S. Wang, Z. Wu, Z. Fei, Q. Cheng, S. Li, and X. Qiu, “MOSS transcribe diarize technical report,” 2026. [Online]. Available: https://arxiv.org/abs/2601.01554

    arXiv 2026

  22. [22]

    SpeakerLM: End-to-end versatile speaker diarization and recognition with multimodal large language models,

    H. Yin, Y . Chen, C. Deng, L. Cheng, H. Wang, C.- H. Tan, Q. Chen, W. Wang, and X. Li, “SpeakerLM: End-to-end versatile speaker diarization and recognition with multimodal large language models,” 2026. [Online]. Available: https://arxiv.org/abs/2508.06372

    arXiv 2026

  23. [23]

    VibeV oice-ASR technical report,

    Z. Peng, J. Yu, Y . Chang, Z. Wang, L. Dong, Y . Hao, Y . Tu, C. Yang, W. Wang, S. Xu, Y . Sun, H. Bao, W. Xu, Y . Zhu, Z. Wang, T. Song, Y . Xia, Z. Chi, S. Huang, L. Wang, C. Ding, S. Wang, X. Chen, and F. Wei, “VibeV oice-ASR technical report,”

  24. [24]

    Available: https://arxiv.org/abs/2601.18184

    [Online]. Available: https://arxiv.org/abs/2601.18184

    arXiv

  25. [25]

    FunAudioLLM: V oice understanding and genera- tion foundation models for natural interaction between humans and llms,

    K. An, Q. Chen, C. Deng, Z. Du, C. Gao, Z. Gao, Y . Gu, T. He, H. Hu, K. Hu, S. Ji, Y . Li, Z. Li, H. Lu, H. Luo, X. Lv, B. Ma, Z. Ma, C. Ni, C. Song, J. Shi, X. Shi, H. Wang, W. Wang, Y . Wang, Z. Xiao, Z. Yan, Y . Yang, B. Zhang, Q. Zhang, S. Zhang, N. Zhao, and S. Zheng, “FunAudioLLM: V oice understanding and genera- tion foundation models for natural ...

    arXiv 2024

  26. [26]

    CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking,

    H. Wang, S. Zheng, Y . Chen, L. Cheng, and Q. Chen, “CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking,” inInterspeech 2023, 2023, pp. 5301– 5305

    2023

  27. [27]

    LoRA: Low-rank adaptation of large lan- guage models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large lan- guage models,” inInternational Conference on Learning Repre- sentations, 2022

    2022

  28. [28]

    Convolutional neural network for speaker change detection in telephone speaker diarization system,

    M. Hruz and Z. Zajic, “Convolutional neural network for speaker change detection in telephone speaker diarization system,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 4945–4949

    2017

  29. [29]

    CHiME-6 challenge: Tackling multispeaker speech recognition for unseg- mented recordings,

    S. Watanabe, M. Mandel, J. Barker, and E. Vincent, “CHiME-6 challenge: Tackling multispeaker speech recognition for unseg- mented recordings,”ArXiv, vol. abs/2004.09249, 2020

    arXiv 2004

  30. [30]

    WenetSpeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,

    B. Zhang, H. Lv, P. Guo, Q. Shao, C. Yang, L. Xie, X. Xu, H. Bu, X. Chen, C. Zeng, D. Wu, and Z. Peng, “WenetSpeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” in International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2022

    2022

  31. [31]

    M2MeT: The ICASSP 2022 multi-channel multi-party meeting transcription challenge,

    F. Yu, S. Zhang, Y . Fu, L. Xie, S. Zheng, Z. Du, W. Huang, P. Guo, Z. Yan, B. Ma, X. Xu, and H. Bu, “M2MeT: The ICASSP 2022 multi-channel multi-party meeting transcription challenge,” inProc. ICASSP. IEEE, 2022

    2022

  32. [32]

    AISHELL-4: An open source dataset for speech enhance- ment, separation, recognition and speaker diarization in confer- ence scenario,

    F. Yihui, C. Luyao, L. Shubo, J. Yukai, K. Yuxiang, C. Zhuo, H. Yanxin, X. Lei, W. Jian, B. Hui, X. Xin, D. Jun, and C. Jing- dong, “AISHELL-4: An open source dataset for speech enhance- ment, separation, recognition and speaker diarization in confer- ence scenario,” inInterspeech, 2021

    2021

  33. [33]

    Qwen2.5: A party of foundation models,

    Q. Team, “Qwen2.5: A party of foundation models,” September

  34. [34]

    Available: https://qwenlm.github.io/blog/qwen2.5

    [Online]. Available: https://qwenlm.github.io/blog/qwen2.5

  35. [35]

    3D-Speaker- Toolkit: An open source toolkit for multi-modal speaker verifi- cation and diarization,

    Y . Chen, S. Zheng, H. Wang, L. Chenget al., “3D-Speaker- Toolkit: An open source toolkit for multi-modal speaker verifi- cation and diarization,” 2025

    2025

  36. [36]

    Efficient and generalizable speaker diarization via structured pruning of self-supervised models,

    J. Han, P. P ´alka, M. Delcroix, F. Landini, J. Rohdin, J. Cernock`y, and L. Burget, “Efficient and generalizable speaker diarization via structured pruning of self-supervised models,”arXiv preprint arXiv:2506.18623, 2025

    arXiv 2025

  37. [37]

    Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,

    Z. Gao, S. Zhang, I. McLoughlin, and Z. Yan, “Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to- end speech recognition,” inProc. Interspeech 2022, 2022, pp. 2063–2067

    2022