Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Enhancing Generalization of Speech Large Language Models with Multi-Task Behavior Imitation and Speech-Text Interleaving

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Training a speech LLM to imitate a frozen text LLM's answers to paired transcripts yields stronger prompt and task generalization, the paper argues.

desk verdict A plausible multi-task behavior-imitation recipe, but the generalization claims rest on a TTS-confounded, GPT-4o-judged benchmark. read the letter →

arxiv 2505.18644 v1 pith:77C5CWBJ submitted 2025-05-24 eess.AS cs.CLcs.SD

Jingran Xie , Xiang Li , Hui Wang , Yue Yu , Yang Xiang , Xixin Wu , Zhiyong Wu This is my paper · ORCID
classification eess.AScs.CLcs.SD
keywords speechlargelanguagemodelbehaviorimitationspeech-textinterleavingzero-shotgeneralizationprompttaskmulti-tasklearningalignment
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

The paper's central claim is that a speech large language model trained to emit the same response as a frozen text LLM would give to the paired transcript becomes more generalized than models trained with supervised annotations on fixed tasks. The proposed MTBI method needs only paired speech and transcripts, uses multi-task behavior imitation plus speech-text interleaving, and zero-shot outperforms current speech LLMs on prompt and task generalization while using less supervised speech data. The authors argue this works because speech is pushed into the text embedding space of the LLM, letting the model inherit the text model's emergent abilities. A reader would care because data annotation is the main bottleneck in making speech models generalize to open-ended tasks.

What carries the argument

The central object is the behavior-imitation training objective: the SLLM is trained to predict the text response that a frozen text LLM generates from the transcript, so the speech representation is optimized to behave like text. The supporting mechanism is speech-text interleaving, where a segment of the transcript is replaced by its synthesized speech and the model must produce the same answer from the mixed input. The trainable connector, a CNN-based subsampler, maps WavLM speech features into the LLM's embedding space, while the speech encoder and the LLM decoder stay frozen.

What would settle it

Re-run the prompt and task generalization tests with recordings of human speakers, not TTS output, on the same prompts and content; if the 85 versus 3 prompt-generalization gap and the 75.3 versus 55.7 role-inference gap shrink or reverse, the central claim fails for natural speech.

Watch

Extended reading notes

Core claim

The authors introduce two-stage 'behavior imitation' alignment. In the first stage, a frozen LLaMA2-7B-chat produces responses from task prompts and transcripts; in the second, the speech model is trained to predict those same responses from the corresponding speech, or from interleaved speech-text input. The model is never trained on task-specific human annotations beyond transcripts, and the interleaving randomly replaces 40–60% of a transcript with a synthesized speech segment so the connector learns to fuse modalities rather than relying on distributional cues. On their benchmark, the configured model with constructed tasks reaches prompt generalization 85, GSM8K 20.1, and speaker-role inference 75.3, versus Qwen2-audio-instruct's 3, 10.5, and 55.7, with comparable ASR word error rate.

Load-bearing premise

The load-bearing assumption is that gains measured on synthesized speech, from CosyVoice 2, transfer to real speech acoustics; if not, the reported margin over other SLLMs may not survive in practice.

Editorial extensions

If this is right

  • Only paired speech and transcripts are needed for training, so large ASR-available corpora can be turned into generalization training data without task-specific annotation.
  • Emergent abilities of the frozen text LLM, including one-shot prompting, transfer to speech input without additional speech supervision.
  • Multi-task behavior imitation beats single-task ASR fine-tuning on prompt generalization, even when trained on the same datasets.
  • The proposed benchmark gives a repeatable way to compare SLLMs on prompt and task generalization rather than only on ASR quality.

Reading between the lines

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

  • If the reported advantage transfers to natural speech, the main scaling bottleneck for speech LLMs shifts from annotated task data to transcript-quality speech data; the paper does not test this transfer directly.
  • Because the prompt-generalization test uses GPT-4o both to generate prompts and to judge compliance, a model that happens to match GPT-4o's formatting preferences could be favored independently of true instruction following.
  • The same two-stage imitation idea could extend beyond speech to other modalities, such as images or video, where a frozen text-only model can generate labels from captions; the paper only hints at this direction.
  • The TTS-based interleaving makes the model dependent on synthetic speech distribution, so evaluating on real human speech could reveal different scaling behavior than the reported numbers suggest.
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 MTBI, a training method for speech large language models (SLLMs) that aligns speech with a frozen text LLM by 'behavior imitation': the text LLM first generates responses from transcripts for multiple tasks, and the SLLM is trained to produce equivalent responses from the corresponding speech. A speech-text interleaving technique, which replaces a 40–60% contiguous segment of the transcript with synthetic speech from CosyVoice 2, is introduced to improve alignment efficiency. The authors construct a benchmark for prompt generalization (10 clips, 100 GPT-4o-generated prompts, judged by GPT-4o) and task generalization (60 CosyVoice 2-synthesized speaker-role utterances and 1,100 CosyVoice 2-synthesized GSM8K problems). They report that MTBI trained on LibriSpeech with three constructed tasks outperforms existing SLLMs such as Qwen2-audio-instruct on prompt and task generalization while using less supervised speech data.

Significance. If validated, the core idea is valuable: it offers a low-cost, transcript-only training signal that could let SLLMs inherit more of a frozen text LLM's generalization ability, and the interleaving mechanism is a plausible way to improve speech-text alignment without time-aligned data. The paper's strengths include a clear architecture, use of public datasets, a useful cascaded-system upper bound (LLaMA2 + ASR), and an ablation study isolating the contributions of ASR SFT, multi-task learning, and interleaving. However, the central empirical claim of superior generalization is not yet convincingly established because the task-generalization test speech is generated by the same TTS engine used in training interleaving, and the prompt-generalization metric rests on a very small, GPT-4o-judged sample without human agreement or statistical confidence. These issues are load-bearing for the abstract's claim of 'a more generalized SLLM' and the table-level comparison with SOTA SLLMs, so the paper needs substantial additional evaluation before the claim can be accepted.

major comments (3)
  1. [§2.3, §3.2, Table 1] The task-generalization evaluation is confounded by a training/test TTS match. The interleaving training in §2.3 synthesizes speech segments with CosyVoice 2, and the entire task-generalization test set in §3.2 (60 speaker-role utterances and 1,100 GSM8K problems) is also synthesized with CosyVoice 2. MTBI therefore receives direct training exposure to CosyVoice 2 acoustics through the connector, whereas comparison models such as Qwen2-audio-instruct were trained on other or more varied speech data. The large advantages in Table 1 for GSM8K (20.1 vs. 10.5) and speaker role (75.3 vs. 55.7) may partly reflect adaptation to CosyVoice 2's acoustic characteristics rather than generalizable alignment. To support the claim that the method 'achieves a more generalized SLLM' under real acoustic conditions, the authors should evaluate on natural human speech for these tasks, or at minimum on a different TTS engine, and report whether the model transfers across acoustic conditions.
  2. [§3.1, Table 1] The prompt-generalization metric is too fragile to support the headline ranking. It uses only 10 clips and 100 prompts generated by GPT-4o, with GPT-4o also judging whether the response 'follows the prompt correctly.' No human agreement, confidence intervals, or significance tests are reported, and the table itself flags that the ASR SFT model's 97% score is artificially high due to prompt-overfitting behavior. With such a small sample and a single judge model, the reported differences (e.g., 85 for Ours w/ Constructed Tasks vs. 3 for Qwen2-audio-instruct) could be inflated by the judge's preferences or by prompt-instance variance. The authors should enlarge the prompt set, report human-evaluated agreement, and provide error bars or significance tests for the prompt-generalization comparison.
  3. [§4.2, Table 1] The comparison with SOTA SLLMs is not controlled for training data domain or architecture, which weakens the 'outperforms SOTA' conclusion beyond the TTS confound. MTBI is trained exclusively on LibriSpeech and constructed tasks, while Qwen2-audio and Qwen2-audio-instruct use substantially larger and more diverse speech corpora. The paper acknowledges this in §4.3, but the conclusion 'outperforms SOTA SLLMs' is nevertheless drawn from a comparison where the baselines were not adapted to the same acoustic distribution. At minimum, the authors should report results for the baselines on a matched natural-speech test set and discuss how the training-domain mismatch affects the interpretation.
minor comments (4)
  1. [§2.2.1] The text says 'our MBTI' but the method is consistently named MTBI; this typo should be corrected.
  2. [§3.2] The name 'CosyV oice 2' contains an erroneous space in two places; it should read 'CosyVoice 2'.
  3. [Table 2] The header row contains stray quote characters ('C, ' 'R, ' 'S, ' and 'I'), which should be cleaned up for clarity.
  4. [§4.1] The paper does not state whether code, trained models, or the constructed benchmark will be released; a data/code availability statement would improve reproducibility, especially since the benchmark is newly introduced.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the paper's behavior-imitation targets come from a frozen text LLM, and its generalization claims are empirically benchmarked, not derived from its inputs; only a minor non-load-bearing self-citation appears.

full rationale

MTBI is an empirical method paper, not a derivation. In the behavior-imitation training (Section 2.2.1), the frozen text LLM first generates responses from transcripts, and then the SLLM is trained to predict those responses from the paired speech; the target is therefore not constructed from the speech input or from the test metrics. The generalization claims are evaluated on held-out prompt-following (Section 3.1) and zero-shot tasks (Section 3.2) that are not among the training objectives, so the reported gains are not forced by the training loss or by any fitted parameter renamed as a prediction. The shared use of CosyVoice 2 for interleaved training segments (Section 2.3) and for test-session synthesis (Section 3.2) is a legitimate external-validity concern about distributional overlap, but it does not make any result equivalent to its input by construction, and no equation-level circularity is exhibited. The only self-citation is reference [17], cited as part of a general list of vanilla speech-LLM frameworks; it is not load-bearing. Accordingly, no circular step can be quoted, and the score reflects only that minor self-citation.

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

The method relies on standard speech-processing components (WavLM, LLaMA2) and introduces no new physical or mathematical entities. The principal assumptions are about the quality of the teacher LLM outputs and the validity of TTS-synthesized speech as a stand-in for real speech in both training interleaving and evaluation.

free parameters (4)
  • interleaving_probability = 0.4
    Probability per batch of applying speech-text interleaving; chosen by hand (Section 2.3).
  • interleaved_segment_ratio = 0.4-0.6
    Random contiguous fraction of the transcription replaced by synthesized speech; chosen by hand (Section 2.3).
  • training_epochs = 1
    Single epoch over LibriSpeech 960h and task data; chosen without reported tuning (Section 4.1).
  • inference_sampling = temperature 0.7, top-p 0.85, max_tokens 100/200
    Sampling hyperparameters selected for inference; not fitted to the benchmark, but affect results (Section 4.1).
assumptions (4)
  • domain assumption WavLM Large last-layer representations contain sufficient linguistic information for the SLLM tasks.
    The frozen encoder is used without verification on these specific tasks; standard in prior SLLM work but assumed here (Section 2.1).
  • domain assumption LLaMA2-7B-chat generates high-quality, sufficiently diverse target responses for continuation, rewriting, and selecting.
    Stage-1 targets are generated by this frozen LLM; if the targets are noisy or repetitive, behavior imitation inherits those flaws (Section 2.2.1).
  • domain assumption TTS-synthesized speech in interleaving is a valid proxy for real speech alignment.
    Interleaving uses CosyVoice 2 output; the paper provides no evidence that this transfers to natural speech at test time (Section 2.3).
  • domain assumption The CosyVoice 2 synthesized test speech is representative of natural speech for task generalization.
    Both test sets (speaker role, GSM8K) are TTS-generated with the same engine, so the evaluation may not measure robustness to real acoustic conditions (Section 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Generalization of Speech Large Language Models with Multi-Task Behavior Imitation and Speech-Text Interleaving." pith.science (2026). https://pith.science/paper/77C5CWBJ

@misc{pith2026250518644,
  author       = {Pith},
  title        = {Pith review of: Enhancing Generalization of Speech Large Language Models with Multi-Task Behavior Imitation and Speech-Text Interleaving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/77C5CWBJ}},
  note         = {Machine review of arXiv:2505.18644}
}
read the original abstract

Large language models (LLMs) have shown remarkable generalization across tasks, leading to increased interest in integrating speech with LLMs. These speech LLMs (SLLMs) typically use supervised fine-tuning to align speech with text-based LLMs. However, the lack of annotated speech data across a wide range of tasks hinders alignment efficiency, resulting in poor generalization. To address these issues, we propose a novel multi-task 'behavior imitation' method with speech-text interleaving, called MTBI, which relies solely on paired speech and transcripts. By ensuring the LLM decoder generates equivalent responses to paired speech and text, we achieve a more generalized SLLM. Interleaving is used to further enhance alignment efficiency. We introduce a simple benchmark to evaluate prompt and task generalization across different models. Experimental results demonstrate that our MTBI outperforms SOTA SLLMs on both prompt and task generalization, while requiring less supervised speech data.

Figures

Figures reproduced from arXiv: 2505.18644 by the authors.

Figure 1
Figure 1. Overview of the SLLM architecture. The training process is conducted in two stages using the same text LLM. In stage 1, we use the LLM to generate responses based on the task prompts and transcripts of speech data. Then, we train the SLLM model with behavior imitation that use the same task prompt and corresponding speech (or interleaved speech) to predict the generated response of the first stage. In stage 2, we on… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 8 canonical work pages

  1. [1]

    Introduction In recent years, large language models (LLMs) [1, 2] have demonstrated remarkable capabilities, particularly in zero-shot scenarios, showcasing their ability to generalize across a wide range of tasks. To leverage this ability for speech, substantial progress [3, 4, 5, 6, 7] has been made in the development of speech large language models (SL...

    work page

  2. [2]

    Hope you<Interleaved Speech>

    Methodology 2.1. Model Architecture We train MTBI on a simple and popular vanilla speech LLM framework [5, 6, 17] as shown in Figure 1, right part. This archi- tecture integrates an frozen speech encoder, a trainable modality connector and a frozen text LLM to generate response. arXiv:2505.18644v1 [eess.AS] 24 May 2025 Figure 1:Overview of the SLLM archit...

    arXiv 2025

  3. [3]

    To address this gap, we draw inspiration from estab- lished evaluation methods in NLP [15, 16]

    Test Set Current research on generalization for SLLMs is relatively scarce. To address this gap, we draw inspiration from estab- lished evaluation methods in NLP [15, 16]. We construct a pre- liminary test set to assess model’s generalization across two di- mensions: prompt generalization and task generalization. 3.1. Prompt Generalization We assess the m...

    work page

  4. [4]

    This task evaluates both the instruction following and reasoning abilities of SLLMs

    Models were prompted to select one role from five options based on the content of the speech. This task evaluates both the instruction following and reasoning abilities of SLLMs. The performance is measured by inference accuracy. GSM8KFor the mathematical reasoning task, we assess the model’s ability to solve arithmetic and reasoning problems presented in...

    work page

  5. [5]

    Experiment 4.1. Experimental Setup For multi-task learning, we use several open-source supervised datasets: LibriSpeech [25] for ASR, Audio Snips [26] for IC, Fluent Speech Commands [27] for SF, and CoV oST 2 [28, 29] for ST. We balance these datasets to ensure a similar proportion across tasks. To maintain a fair comparison, our MTBI uses the same datase...

    work page

  6. [6]

    Conclusion In this work, we propose a novel approach to enhance the gener- alization capabilities of SLLMs through multi-task behavior im- itation with speech-text interleaving. Our experiments demon- strate that our method, trained solely on ASR-supervised data using three constructed tasks, significantly boosts the general- ization on both prompts and t...

    work page

  7. [7]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Ale- man, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  8. [8]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    arXiv 2023

Show all 36 references
  1. [9]

    Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,”arXiv preprint arXiv:2305.11000, 2023

    2023 arXiv

  2. [10]

    Llasm: Large language and speech model,

    Y . Shu, S. Dong, G. Chen, W. Huang, R. Zhang, D. Shi, Q. Xiang, and Y . Shi, “Llasm: Large language and speech model,”arXiv preprint arXiv:2308.15930, 2023

    2023 arXiv

  3. [11]

    Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    2023 arXiv

  4. [12]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Linet al., “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    2024 arXiv

  5. [13]

    Mini-omni: Language models can hear, talk while thinking in streaming,

    Z. Xie and C. Wu, “Mini-omni: Language models can hear, talk while thinking in streaming,”arXiv preprint arXiv:2408.16725, 2024

    2024 arXiv

  6. [14]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog, vol. 1, no. 8, p. 9, 2019

    2019

  7. [15]

    Multi-task learning in natu- ral language processing: An overview,

    S. Chen, Y . Zhang, and Q. Yang, “Multi-task learning in natu- ral language processing: An overview,”ACM Computing Surveys, vol. 56, no. 12, pp. 1–32, 2024

    2024

  8. [16]

    Blsp: Bootstrapping language-speech pre-training via behavior alignment of continuation writing,

    C. Wang, M. Liao, Z. Huang, J. Lu, J. Wu, Y . Liu, C. Zong, and J. Zhang, “Blsp: Bootstrapping language-speech pre-training via behavior alignment of continuation writing,”arXiv preprint arXiv:2309.00916, 2023

    2023 arXiv

  9. [17]

    Speechverse: A large-scale generalizable audio language model,

    N. Das, S. Dingliwal, S. Ronanki, R. Paturi, Z. Huang, P. Mathur, J. Yuan, D. Bekal, X. Niu, S. M. Jayanthiet al., “Speechverse: A large-scale generalizable audio language model,”arXiv preprint arXiv:2405.08295, 2024

    2024 arXiv

  10. [18]

    Scaling speech-text pre-training with synthetic interleaved data,

    A. Zeng, Z. Du, M. Liu, L. Zhang, S. Jiang, Y . Dong, and J. Tang, “Scaling speech-text pre-training with synthetic interleaved data,” arXiv preprint arXiv:2411.17607, 2024

    2024 arXiv

  11. [19]

    Spirit-lm: Interleaved spoken and written language model,

    T. A. Nguyen, B. Muller, B. Yu, M. R. Costa-Jussa, M. El- bayad, S. Popuri, P.-A. Duquenne, R. Algayres, R. Mavlyutov, I. Gatet al., “Spirit-lm: Interleaved spoken and written language model,”arXiv preprint arXiv:2402.05755, 2024

    2024 arXiv

  12. [20]

    Gpteval: A survey on assessments of chatgpt and gpt-4,

    R. Mao, G. Chen, X. Zhang, F. Guerin, and E. Cambria, “Gpteval: A survey on assessments of chatgpt and gpt-4,”arXiv preprint arXiv:2308.12488, 2023

    2023 arXiv

  13. [21]

    Instruction-following evaluation for large language models,

    J. Zhou, T. Lu, S. Mishra, S. Brahma, S. Basu, Y . Luan, D. Zhou, and L. Hou, “Instruction-following evaluation for large language models,”arXiv preprint arXiv:2311.07911, 2023

    2023 arXiv

  14. [22]

    Llm-based nlg evaluation: Current status and challenges,

    M. Gao, X. Hu, J. Ruan, X. Pu, and X. Wan, “Llm-based nlg evaluation: Current status and challenges,”arXiv preprint arXiv:2402.01383, 2024

    2024 arXiv

  15. [23]

    Leveraging chain of thought towards empathetic spoken dialogue without corresponding question-answering data,

    J. Xie, S. Lei, Y . Yu, Y . Xiang, H. Wang, X. Wu, and Z. Wu, “Leveraging chain of thought towards empathetic spoken dialogue without corresponding question-answering data,”arXiv preprint arXiv:2501.10937, 2025

    2025 arXiv

  16. [24]

    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. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  17. [25]

    Llama-adapter: Efficient fine-tuning of language models with zero-init attention,

    R. Zhang, J. Han, C. Liu, P. Gao, A. Zhou, X. Hu, S. Yan, P. Lu, H. Li, and Y . Qiao, “Llama-adapter: Efficient fine-tuning of language models with zero-init attention,”arXiv preprint arXiv:2303.16199, 2023

    2023 arXiv

  18. [26]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” inInternational conference on machine learning. PMLR, 2023, pp. 19 730–19 742

    2023

  19. [27]

    Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wanget al., “Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,”arXiv preprint arXiv:2412.10117, 2024

    2024 arXiv

  20. [28]

    Hello gpt-4o,

    OpenAI, “Hello gpt-4o,” OpenAI, Tech. Rep., 2024. [Online]. Available: https://openai.com/index/hello-gpt-4o

    2024

  21. [29]

    Train- ing verifiers to solve math word problems,

    K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakanoet al., “Train- ing verifiers to solve math word problems,”arXiv preprint arXiv:2110.14168, 2021

    2021 arXiv

  22. [30]

    Measuring mathemat- ical problem solving with the math dataset,

    D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song, and J. Steinhardt, “Measuring mathemat- ical problem solving with the math dataset,”arXiv preprint arXiv:2103.03874, 2021

    2021 arXiv

  23. [31]

    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

  24. [32]

    Snips voice platform: an embedded spoken language under- standing system for private-by-design voice interfaces,

    A. Coucke, A. Saade, A. Ball, T. Bluche, A. Caulier, D. Leroy, C. Doumouro, T. Gisselbrecht, F. Caltagirone, T. Lavrilet al., “Snips voice platform: an embedded spoken language under- standing system for private-by-design voice interfaces,”arXiv preprint arXiv:1805.10190, 2018

    2018 arXiv

  25. [33]

    Speech model pre-training for end-to-end spoken language understanding,

    L. Lugosch, M. Ravanelli, P. Ignoto, V . S. Tomar, and Y . Ben- gio, “Speech model pre-training for end-to-end spoken language understanding,”arXiv preprint arXiv:1904.03670, 2019

    1904 arXiv

  26. [34]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,”arXiv preprint arXiv:1912.06670, 2019

    1912 arXiv

  27. [35]

    Covost 2 and massively multilingual speech translation

    C. Wang, A. Wu, J. Gu, and J. Pino, “Covost 2 and massively multilingual speech translation.” inInterspeech, 2021, pp. 2247– 2251

    2021

  28. [36]

    Robust speech recognition via large- scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large- scale weak supervision,” 2022. [Online]. Available: https: //arxiv.org/abs/2212.04356

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.