Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Customizing Speech Recognition Model with Large Language Model Feedback

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 An LLM's text scores can reward ASR hypotheses and cut entity error rate by 21% over self-training.

desk verdict A genuinely new use of LLM log-probabilities as RL rewards for unsupervised ASR domain adaptation, but the load-bearing assumption that those scores rank hypotheses by transcription quality is never tested. read the letter →

arxiv 2506.11091 v2 pith:GZJ2YBCP submitted 2025-06-05 cs.CL cs.SDeess.AS

Shaoshi Ling , Guoli Ye This is my paper
classification cs.CLcs.SDeess.AS
keywords unsuperviseddomainadaptationautomaticspeechrecognitionlargelanguagemodelfeedbackreinforcementlearningnamedentitycontextualcustomizationself-trainingreward
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 proposes a fully unsupervised way to adapt a speech recognition model to a new domain: instead of human transcriptions, a large language model supplies the training signal. The ASR model samples several candidate transcriptions for each unlabeled utterance, an LLM scores each candidate by its summed log-probability under a context prompt such as "Generate a message for ordering at Starbucks," and reinforcement learning pushes the ASR model toward high-scoring hypotheses. Across three domain-mismatch test sets—Starbucks ordering, voice commands, and display-friendly numbers—the method beats standard self-training, with an average 8% relative WER reduction and 21% relative entity-WER improvement. If this holds, domain customization of speech recognizers can be driven by text-only feedback and a small amount of unlabeled audio, without expensive annotation.

What carries the argument

The load-bearing object is the reward function in Eq. (1), which turns a text-only LLM into a critic for speech hypotheses. For each sampled transcription, the LLM computes the sum of log-probabilities over the hypothesis tokens when prompted to generate a message matching the target domain ("Generate a voice command message," for example), and this is added to the ASR model's acoustic score. The resulting reward is dense—every hypothesis gets a score—and the same reward feeds all three fine-tuning objectives: RAFT keeps only the best hypothesis, DPO forms preference pairs from best and worst, and GRPO normalizes rewards within a group and optimizes a clipped policy ratio with a KL penalty. The comparison across algorithms shows that the reward signal, not the specific RL objective, is what drives the improvement.

What would settle it

Take a set of ASR candidate hypotheses with known reference transcripts, compute the Eq. (1) reward for each and its WER/EWER, and check whether the reward orders the hypotheses correctly; a candidate with a wrong but fluent entity spelling that scores above the correct transcription would indicate the reward is misdirected and the reported gains should not generalize beyond the test sets.

Watch

Extended reading notes

Core claim

The central claim is that a frozen text LLM can act as a reliable reward model for ASR adaptation. Formally, the reward for a hypothesis $a_i$ given audio $X_i$ and a contextual prompt $CP$ is $r_{i,n}=P_{\mathrm{llm}}(a_{i,n}|CP)+\lambda P_{\mathrm{asr}}(a_{i,n}|X_i)$, the sum of the LLM's log-probability of the hypothesis and the ASR model's own score. With this reward, RAFT, DPO, and GRPO all improve over the baseline and over self-training; the best system reports an average 15% relative WER and 50% relative EWER improvement over the unadapted baseline, and 8% and 21% respectively over self-training. The same pipeline also improves formatted readability on SPGISpeech, with 7% relative WER and 5.5% relative TER reduction. The authors interpret the gains as evidence that LLM contextual knowledge transfers to the ASR model through RL, particularly for rare entities.

Load-bearing premise

The load-bearing premise is that the LLM's summed log-probability of a candidate transcription, conditioned on a hand-written prompt, is a trustworthy measure of transcription quality—including the spelling of rare named entities—so that maximizing it genuinely reduces recognition errors.

Editorial extensions

If this is right

  • Unlabeled audio plus a one-line prompt is enough to move an ASR model into a new domain; target-domain transcriptions are not required.
  • Entity-heavy applications stand to gain most: the method reports a 21% relative entity-WER improvement over self-training on average.
  • The adapted model needs no LLM at inference time, so the domain customization does not add per-query cost.
  • The approach is complementary to LLM post-processing and could be stacked with rescoring or rewriting for further gains.
  • All three RL objectives perform similarly, so practitioners can choose the simplest (RAFT) without sacrificing accuracy.

Reading between the lines

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

  • A direct check of the reward's validity—correlating LLM scores against WER on held-out hypotheses—would tell whether the 21% EWER gain comes from the LLM preferring correct entities or merely preferring fluent text.
  • The 300-utterance adaptation sets suggest the method may work at very low data volumes; a data-scaling study would reveal the minimum number of unlabeled utterances needed.
  • Because the reward is computed in text space, the same framework should transfer to any sequence model, including streaming ASR or other modalities, provided the model can sample diverse hypotheses.
  • The context prompt itself behaves like a hyper-parameter; the ablation shows context changes EWER by 5%, so prompt engineering could yield further gains or, if done poorly, misdirect the adaptation.
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 an unsupervised domain adaptation method for automatic speech recognition in which a text LLM serves as a reward model for RL-based fine-tuning of the ASR policy. For each unlabeled audio utterance, the ASR model samples n candidate hypotheses; the reward is the sum of the LLM's log-probability of the hypothesis given a hand-written domain prompt plus a weighted ASR acoustic score (Eq. 1). The method is evaluated with RAFT, DPO, and GRPO on three internal domain datasets (food ordering, voice commands, display-format numbers) and on SPGISpeech for readability customization. Compared with a simple self-training baseline, the best variant achieves on average an 8% relative WER reduction and a 21% relative EWER reduction.

Significance. The central idea of using LLM probability scores as dense rewards for ASR domain adaptation is novel and potentially useful, and the paper's experimental scope covers three diverse customization scenarios plus a public-style readability set. The method is clearly described, the RL algorithms are standard, and the comparison against LLM-based post-processing is informative. However, the load-bearing assumption that the LLM score separates correct from incorrect transcriptions for a fixed audio input is not validated, the acoustic weight λ is never reported, and the evaluation lacks statistical significance testing. If the reward-validation concern is addressed, the contribution would be solid; as presented, the current evidence does not conclusively attribute the reported gains to LLM feedback.

major comments (3)
  1. [III-B, Eq. (1)] The reward in Eq. (1) combines an LLM score P_llm(a|CP) that is not conditioned on the audio X with an ASR acoustic score λP_asr(a|X). No evidence is provided that P_llm ranks hypotheses by transcription quality for a fixed utterance: there is no correlation analysis against WER or EWER, no oracle-reward control, and no failure analysis for entity-bearing utterances. The ablation in Table III only demonstrates that the domain-specific prompt changes the LLM score; it does not show that the LLM score separates correct from incorrect transcriptions. Since λ is never reported, the gains in Table I could partly arise from the ASR self-training term rather than from LLM feedback. Please report λ for each experiment, run an ablation with λ=0, and provide a reward-quality analysis (e.g., an oracle selection upper bound or a correlation of P_llm with reference-based accuracy on a small labeled set).
  2. [IV-C, Table I] The reported improvements (8% relative WER and 21% relative EWER averaged over three datasets) are presented without confidence intervals, significance tests, or multiple seeds, and the test sets are small (on the order of 200–800 utterances). The abstract's phrasing "up to a 21% improvement" is imprecise because 21% is the average EWER gain, not a maximum. Please add bootstrap confidence intervals or per-seed standard deviations, and report the exact test-set sizes and per-dataset uncertainty.
  3. [IV-C] The self-training baseline selects the hypothesis with the highest ASR score for fine-tuning, which is a high-confidence selection without thresholding or pseudo-label denoising. Contemporary unsupervised domain adaptation baselines include confidence-filtered pseudo-labeling and uncertainty-based methods (e.g., references [25] and [28]). To support the claim that LLM feedback outperforms conventional self-training, please compare against a stronger confidence-filtered baseline, or justify why the simple high-score baseline is the appropriate reference for this comparison.
minor comments (4)
  1. [Table I] Several table cells have concatenated numbers without separators (e.g., "12.5917.42" and "12.4917.81"); these should be formatted as distinct entries to avoid ambiguity.
  2. [III-C, Eq. (3)] The DPO objective assumes a fixed reference model π_ref; the paper should state explicitly that π_ref is the frozen pretrained ASR model throughout training, and clarify how on-policy sampling interacts with the off-policy derivation of DPO.
  3. [IV-D] The definition of token error rate (TER) as a space-separated, case-sensitive token error rate is non-standard; please rename it or define it explicitly as a case-sensitive token-level error rate to avoid confusion with translation edit rate.
  4. [I] The introduction states that the proposed reward "avoids common issues such as reward hacking," but no evidence or analysis supporting this claim is provided; please soften the claim or add a discussion of potential reward hacking and why the combined reward mitigates it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported WER/EWER gains are externally measured, and the LLM reward is an independent signal.

full rationale

The paper's central claim is that LLM log-probability feedback improves ASR domain adaptation, evaluated by WER and EWER on held-out test sets. The reward in Eq. (1) combines an external LLM score P_llm(a|CP) with the ASR model's own likelihood P_asr(a|X). Although including the ASR likelihood can bias sample selection toward the model's prior, this is a potential optimization concern, not a circular derivation: the reported improvements are not constructed from the reward, and no parameter is fitted to the test metric. The LLM score is an external signal, and the ablation in Table III shows the contextual prompt changes outcomes, indicating the LLM term carries information. The only self-citation (ref [27]) appears in the related-work enumeration of pseudo-label denoising techniques and is not load-bearing. No step in the derivation reduces by definition to its own input, so no circularity is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central method rests on the unvalidated assumption that LLM text log-probabilities, conditioned on a short prompt, are a good reward for ASR transcription quality. The linear combination with the ASR model's own score adds a tunable parameter that is never reported. No new entities are introduced, but the prompt and reward design are ad hoc choices.

free parameters (1)
  • lambda (acoustic score weight) = not reported
    Eq. (1) combines LLM log-probabilities with ASR acoustic scores using a tunable scalar lambda. The paper never specifies its value or how it was chosen, yet all reward computations and downstream results depend on it.
assumptions (4)
  • domain assumption LLM log-probability score P_llm(a|CP) is a valid reward for transcription quality, including named entities.
    Introduced in Section III-B, Eq. (1); no evidence in the paper that this score correlates with transcription correctness or entity accuracy.
  • ad hoc to paper The hand-written prompt 'Generate a message optimized for {CP}' elicits scores that favor accurate transcriptions over fluent but incorrect text.
    Prompt design is specific to this work and is not validated against alternative prompts or against transcriptions.
  • ad hoc to paper The acoustic model score P_asr(a|X) can be linearly combined with the LLM score via a single scalar lambda without calibration.
    Eq. (1); lambda is never specified and the two score distributions are not normalized or calibrated.
  • domain assumption Unlabeled adaptation audio (300 utterances per domain) is representative of the test domain.
    Section IV-C; no coverage analysis is provided, and test sets are up to 800 utterances, so sampling bias is possible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Customizing Speech Recognition Model with Large Language Model Feedback." pith.science (2026). https://pith.science/paper/GZJ2YBCP

@misc{pith2026250611091,
  author       = {Pith},
  title        = {Pith review of: Customizing Speech Recognition Model with Large Language Model Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZJ2YBCP}},
  note         = {Machine review of arXiv:2506.11091}
}
read the original abstract

Automatic speech recognition (ASR) systems have achieved strong performance on general transcription tasks. However, they continue to struggle with recognizing rare named entities and adapting to domain mismatches. In contrast, large language models (LLMs), trained on massive internet-scale datasets, are often more effective across a wide range of domains. In this work, we propose a reinforcement learning based approach for unsupervised domain adaptation, leveraging unlabeled data to enhance transcription quality, particularly the named entities affected by domain mismatch, through feedback from a LLM. Given contextual information, our framework employs a LLM as the reward model to score the hypotheses from the ASR model. These scores serve as reward signals to fine-tune the ASR model via reinforcement learning. Our method achieves a 21\% improvement on entity word error rate over conventional self-training methods.

Figures

Figures reproduced from arXiv: 2506.11091 by the authors.

Figure 1
Figure 1. The Overview 1) Rejection sampling fine-tuning: We follow the rejection sampling fine-tuning (RAFT) in the literature [33], [34]. In the generating rewards step, we retains only the hypothesis with the highest reward and we only aggregate those positive samples into Dpos to fine-tune the model. The objective function is to maximize the log-likelihood over the dataset Dpos. LRAF T = X a∈Dp logπ(a|X) (2) 2) DPO: The D… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 22 canonical work pages

  1. [25]

    Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,

    S. Khurana, N. Moritz, T. Hori, and J. Le Roux, “Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6553–6557

    work page 2021

  2. [28]

    Self-taught recognizer: Toward unsupervised adaptation for speech foundation models,

    Y . Hu, C. Chen, C.-H. Yang, C. Qin, P.-Y . Chen, E.-S. Chng, and C. Zhang, “Self-taught recognizer: Toward unsupervised adaptation for speech foundation models,”Advances in Neural Information Processing Systems, vol. 37, pp. 29 566–29 594, 2024

    work page 2024

  3. [1]

    Recent advances in end-to-end automatic speech recog- nition,

    J. Liet al., “Recent advances in end-to-end automatic speech recog- nition,”APSIPA Transactions on Signal and Information Processing, vol. 11, no. 1

    work page

  4. [2]

    Adaptation algorithms for neural network-based speech recognition: An overview,

    P. Bell, J. Fainberg, O. Klejch, J. Li, S. Renals, and P. Swietojanski, “Adaptation algorithms for neural network-based speech recognition: An overview,”IEEE Open Journal of Signal Processing, vol. 2, pp. 33–66, 2020

    work page 2020

  5. [3]

    Self-taught learning: transfer learning from unlabeled data,

    R. Raina, A. Battle, H. Lee, B. Packer, and A. Y . Ng, “Self-taught learning: transfer learning from unlabeled data,” inProceedings of the 24th international conference on Machine learning, 2007, pp. 759–766

    work page 2007

  6. [4]

    Confidence estimation for attention-based sequence-to- sequence models for speech recognition,

    Q. Li, D. Qiu, Y . Zhang, B. Li, Y . He, P. C. Woodland, L. Cao, and T. Strohman, “Confidence estimation for attention-based sequence-to- sequence models for speech recognition,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6388–6392

    2021

  7. [5]

    Gpt-4 technical report,

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

    arXiv 2023

  8. [6]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    arXiv 2025

Show all 37 references
  1. [7]

    Phi- 4 technical report,

    M. Abdin, J. Aneja, H. Behl, S. Bubeck, R. Eldan, S. Gunasekar, M. Harrison, R. J. Hewett, M. Javaheripi, P. Kauffmannet al., “Phi- 4 technical report,”arXiv preprint arXiv:2412.08905, 2024

    2024 arXiv

  2. [8]

    Massively multilingual shal- low fusion with large language models,

    K. Hu, T. N. Sainath, B. Li, N. Du, Y . Huang, A. M. Dai, Y . Zhang, R. Cabrera, Z. Chen, and T. Strohman, “Massively multilingual shal- low fusion with large language models,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)...

    2023

  3. [9]

    Prompting large language models for zero-shot domain adaptation in speech recognition,

    Y . Li, Y . Wu, J. Li, and S. Liu, “Prompting large language models for zero-shot domain adaptation in speech recognition,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  4. [10]

    Large- scale language model rescoring on long-form data,

    T. Chen, C. Allauzen, Y . Huang, D. Park, D. Rybach, W. R. Huang, R. Cabrera, K. Audhkhasi, B. Ramabhadran, P. J. Morenoet al., “Large- scale language model rescoring on long-form data,” inICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processi...

    2023

  5. [11]

    Applying llms for rescoring n-best asr hypotheses of casual conversations: Effects of domain adaptation and context carry-over,

    A. Ogawa, N. Kamo, K. Matsuura, T. Ashihara, T. Moriya, T. Kano, N. Tawara, and M. Delcroix, “Applying llms for rescoring n-best asr hypotheses of casual conversations: Effects of domain adaptation and context carry-over,”arXiv preprint arXiv:2406.18972, 2024

    2024 arXiv

  6. [12]

    Can generative large language models perform asr error correction?

    R. Ma, M. Qian, P. Manakul, M. Gales, and K. Knill, “Can generative large language models perform asr error correction?”arXiv preprint arXiv:2307.04172, 2023

    2023 arXiv

  7. [13]

    Denoising lm: Pushing the limits of error correction models for speech recognition,

    Z. Gu, T. Likhomanenko, H. Bai, E. McDermott, R. Collobert, and N. Jaitly, “Denoising lm: Pushing the limits of error correction models for speech recognition,”arXiv preprint arXiv:2405.15216, 2024

    2024

  8. [14]

    Asr error correction using large language models,

    R. Ma, M. Qian, M. Gales, and K. Knill, “Asr error correction using large language models,”IEEE Transactions on Audio, Speech and Language Processing, 2025

    2025

  9. [15]

    Constitutional ai: Harmlessness from ai feedback,

    Y . Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnonet al., “Constitutional ai: Harmlessness from ai feedback,”arXiv preprint arXiv:2212.08073, 2022

    2022 arXiv

  10. [16]

    Direct language model alignment from online ai feedback,

    S. Guo, B. Zhang, T. Liu, T. Liu, M. Khalman, F. Llinares, A. Rame, T. Mesnard, Y . Zhao, B. Piotet al., “Direct language model alignment from online ai feedback,”arXiv preprint arXiv:2402.04792, 2024

    2024 arXiv

  11. [17]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,”Advances in Neural Information Processing Systems, vol. 36, pp. 53 728–53 741, 2023

    2023

  12. [18]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  13. [19]

    Boost- ing cross-domain speech recognition with self-supervision,

    H. Zhu, G. Cheng, J. Wang, W. Hou, P. Zhang, and Y . Yan, “Boost- ing cross-domain speech recognition with self-supervision,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 471–485, 2023

    2023

  14. [20]

    An unsupervised deep domain adaptation approach for robust speech recognition,

    S. Sun, B. Zhang, L. Xie, and Y . Zhang, “An unsupervised deep domain adaptation approach for robust speech recognition,”Neurocomputing, vol. 257, pp. 79–87, 2017

    2017

  15. [21]

    A multi- discriminator cyclegan for unsupervised non-parallel speech domain adaptation,

    E. Hosseini-Asl, Y . Zhou, C. Xiong, and R. Socher, “A multi- discriminator cyclegan for unsupervised non-parallel speech domain adaptation,”arXiv preprint arXiv:1804.00522, 2018

    2018 arXiv

  16. [22]

    Toward domain-invariant speech recognition via large scale training,

    A. Narayanan, A. Misra, K. C. Sim, G. Pundak, A. Tripathi, M. Elfeky, P. Haghani, T. Strohman, and M. Bacchiani, “Toward domain-invariant speech recognition via large scale training,” in2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018, pp. 441–447

    2018

  17. [23]

    Large- scale domain adaptation via teacher-student learning,

    J. Li, M. L. Seltzer, X. Wang, R. Zhao, and Y . Gong, “Large- scale domain adaptation via teacher-student learning,”arXiv preprint arXiv:1708.05466, 2017

    2017 arXiv

  18. [24]

    Conditional teacher-student learning,

    Z. Meng, J. Li, Y . Zhao, and Y . Gong, “Conditional teacher-student learning,” inICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 6445–6449

    2019

  19. [26]

    Large-scale asr domain adaptation using self-and semi-supervised learning,

    D. Hwang, A. Misra, Z. Huo, N. Siddhartha, S. Garg, D. Qiu, K. C. Sim, T. Strohman, F. Beaufays, and Y . He, “Large-scale asr domain adaptation using self-and semi-supervised learning,” inICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processin...

    2022

  20. [27]

    Improving pseudo-label training for end-to-end speech recognition using gradient mask,

    S. Ling, C. Shen, M. Cai, and Z. Ma, “Improving pseudo-label training for end-to-end speech recognition using gradient mask,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8397–8401

    2022

  21. [29]

    Partitioning attention weight: Mitigat- ing adverse effect of incorrect pseudo-labels for self-supervised asr,

    J.-H. Lee and J.-H. Chang, “Partitioning attention weight: Mitigat- ing adverse effect of incorrect pseudo-labels for self-supervised asr,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 891–905, 2023

    2023

  22. [30]

    Self-critical sequence training for automatic speech recognition,

    C. Chen, Y . Hu, N. Hou, X. Qi, H. Zou, and E. S. Chng, “Self-critical sequence training for automatic speech recognition,” inICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 3688–3692

    2022

  23. [31]

    Sequence-to-sequence asr optimization via reinforcement learning,

    A. Tjandra, S. Sakti, and S. Nakamura, “Sequence-to-sequence asr optimization via reinforcement learning,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5829–5833

    2018

  24. [32]

    Reinforcement learning of speech recognition system based on policy gradient and hypothesis selection,

    T. Kala and T. Shinozaki, “Reinforcement learning of speech recognition system based on policy gradient and hypothesis selection,” in2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 5759–5763

    2018

  25. [33]

    A minimalist approach to llm reasoning: from rejection sampling to reinforce,

    W. Xiong, J. Yao, Y . Xu, B. Pang, L. Wang, D. Sahoo, J. Li, N. Jiang, T. Zhang, C. Xionget al., “A minimalist approach to llm reasoning: from rejection sampling to reinforce,”arXiv preprint arXiv:2504.11343, 2025

    2025 arXiv

  26. [34]

    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

    2023 arXiv

  27. [35]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  28. [36]

    A call for clarity in beam search: How it works and when it stops,

    J. Kasai, K. Sakaguchi, R. L. Bras, D. Radev, Y . Choi, and N. A. Smith, “A call for clarity in beam search: How it works and when it stops,” arXiv preprint arXiv:2204.05424, 2022

    2022 arXiv

  29. [37]

    Spgispeech: 5,000 hours of transcribed financial audio for fully formatted end-to-end speech recognition,

    P. K. O’Neill, V . Lavrukhin, S. Majumdar, V . Noroozi, Y . Zhang, O. Kuchaiev, J. Balam, Y . Dovzhenko, K. Freyberg, M. D. Shul- manet al., “Spgispeech: 5,000 hours of transcribed financial audio for fully formatted end-to-end speech recognition,”arXiv preprint arXiv:2104.02014, 2021

    2021 arXiv

Pith tools

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