Pith. sign in

REVIEW 2 major objections 5 minor 43 references

Seewo's Submission to MLC-SLM: Lessons learned from Speech Reasoning Language Models

T0 review · 2 major / 5 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 The paper reports a multilingual ASR system that reaches 11.57% WER/CER on Track 1 and 17.67% tcpWER/tcpCER on Track 2 by training a speech language model to reflect and self-correct before transcribing.

desk verdict Solid challenge paper with strong results, but the CoT ablation is confounded with the loss change; the claimed reasoning benefit is not isolated. read the letter →

arxiv 2506.13300 v3 pith:VGTRBJPD submitted 2025-06-16 cs.CL cs.AIcs.SDeess.AS

Bo Li , Chengben Xu , Wufeng Zhang This is my paper · ORCID
classification cs.CLcs.AIcs.SDeess.AS
keywords multilingualspeechrecognitionlanguagemodelchain-of-thoughtreinforcementlearningwithverifiablerewardsself-correctionspeakerdiarizationcurriculumLoRA
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 argues that a speech language model can be taught to reason about its own transcription errors and correct them, and that this self-correction is what lets a multilingual ASR system beat the challenge baselines by a wide margin. It reports a training recipe in which the model first learns to map speech to text, then learns a structured 'think before transcribing' output, and finally is optimized with reinforcement learning rewards that verify whether the correction actually helped. On the evaluation set the best system reaches 11.57% WER/CER on Track 1 and 17.67% tcpWER/tcpCER on Track 2, against baselines of 20.17% and 60.39%. The paper's practical point is that reasoning-style training, previously applied to text LLMs, transfers to speech, and that outcome-based rewards work where process-based rewards do not.

What carries the argument

The load-bearing object is the staged fine-tuning pipeline built on an encoder-projector-decoder architecture: a Whisper encoder mapped through a learned projector to a Babel-9B-Chat decoder. Four mechanisms carry the argument: (1) a chat template with special tokens (<LANG XX>, <speech>, <transcribe>, <think>) that forces structured output; (2) prompt-loss weighting, which downweights the <think> section so the model learns to transcribe after reflecting; (3) LoRA adaptation of both encoder and decoder; and (4) reinforcement learning with verifiable rewards RF1 (format), RF2 (transcription accuracy), and RF5 (second hypothesis improves over first). The pipeline's order matters: each stage is a prerequisite for the next, and removing the format reward collapses the output structure.

What would settle it

Train the stage-3 model for 3,000 additional steps without the reflective data, using the same loss schedule; if it reaches the same 13.42% error rate as the 10,000-step checkpoint, the reflection data was not the operative ingredient.

Watch

Extended reading notes

Core claim

The central claim is that a staged curriculum—projector alignment, task-token adaptation, LoRA fine-tuning, Chain-of-Thought SFT with a weighted loss, and reinforcement learning with verifiable rewards—lets a Whisper large-v3-turbo encoder and a Babel-9B-Chat decoder produce accurate multilingual transcriptions with explicit reflection. The best reward configuration scores only structure, transcription accuracy, and whether the second hypothesis improves on the first; rewarding the quality of the intermediate reasoning (error-type or error-detail accuracy) hurt or did nothing. The paper therefore concludes that outcome-based verification can induce useful self-correction in speech LMs even when the generated reasoning trace is not itself faithful. It also reports that the CoT SFT stage initially made WER/CER worse, from 15.48% to 35.03%, until the loss was reweighted to focus on the transcribed output rather than the reasoning section.

Load-bearing premise

The claim that reflective training caused the 2-point error drop rests on comparing checkpoints that differ in both training data and training steps, so extra fine-tuning alone could account for part of the gain.

Editorial extensions

If this is right

  • If the pipeline is right, the official challenge baselines are not hard ceilings for end-to-end speech language models; a tuned curriculum plus RLVR can cut WER/CER by more than 8 absolute points on the evaluation set.
  • Outcome-based RLVR (RF1+RF2+RF5) is the configuration to copy; adding process rewards RF3 and RF4 actually degraded WER/CER relative to RF1+RF2.
  • Chain-of-Thought data helps only if the loss is reweighted; naive CoT SFT raised WER/CER from 15.48% to 35.03%.
  • For diarization-plus-ASR, capping the VAD segment at 8 seconds instead of 60 improves tcpWER from 24.64% to 18.03% on the development set, because segments then fit the encoder's limit.
  • Conversational context can improve ASR (14.30% vs 15.48% on the development set), but the paper measured only an upper bound using ground-truth previous utterances.

Reading between the lines

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

  • An untested but natural reading is that the reasoning tokens are a scaffold rather than a cause: if outcome rewards alone drive gains, the model may learn to emit plausible 'thinking' without using it. A clean test would mask or remove the <think> section at inference and compare WER/CER.
  • The CoT-vs-length confound could be settled by training a control from the stage-3 checkpoint for 3,000 more steps without CoT data; if it reaches 13.42% too, the attribution to reflection collapses.
  • The context experiment suggests that end-to-end conversational ASR could gain from utterance history, but realizing it requires a streaming or batched inference pipeline; the paper's own inference inefficiency leaves this untested.
  • Process rewards may fail because error-detail supervision is too sparse and noisy; a denser reward derived from aligning predicted error spans with ground-truth edits would be a testable next step.
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

2 major / 5 minor

Summary. This paper reports Seewo's systems for both tracks of the MLC-SLM challenge: a speech language model for multilingual ASR (Track 1) and a speaker diarization pipeline combined with ASR (Track 2). The ASR system uses a Whisper large-v3-turbo encoder, a Babel-9B-Chat decoder, and a learned projector, trained through a multi-stage pipeline: projector-only training, special-token adaptation, LoRA fine-tuning with language-specific prompts, a Chain-of-Thought (CoT) SFT stage with a prompt-loss-weighting (PLW) objective, and a final Dr. GRPO reinforcement learning stage with several reward functions. The diarization system concatenates voice activity detection, speaker embedding extraction, and spectral clustering. On the official evaluation set, the best system achieves 11.57% WER/CER (Track 1) and 17.67% tcpWER/tcpCER (Track 2), outperforming the official baselines by large margins. The paper claims that each training component, especially CoT-based reflection and RLVR-based self-correction, contributes to the gains.

Significance. If the component-level claims were established, this would be a useful engineering recipe for end-to-end multilingual conversational ASR, with unusually detailed ablations of reward functions, decoding strategies, and diarization configurations. The overall system results are strong and beat the provided baselines by substantial margins, and the paper is commendably transparent about several limitations, including the upper-bound nature of the context experiment and the authors' own observation that intermediate reasoning did not demonstrate a direct causal influence on output. However, the central attribution of improvements to reasoning/CoT is not supported by the experiments as presented, and the reward-function ablations lack any statistical validation. The paper is a solid system description, but its headline claims about reasoning require additional experimentation before they can be accepted.

major comments (2)
  1. [§3.4, Table 1] The claim that CoT SFT improves WER/CER is confounded. The comparison of ckpt-7000 (15.48%) with ckpt-10000 (13.42%) changes three factors simultaneously: 3000 additional training steps, the introduction of CoT data, and the PLW loss (Eq. 2). The internal comparison ckpt-9000 (35.03%) shows that CoT data with the standard causal LM loss makes performance dramatically worse, and the recovery to 13.42% occurs only after the PLW loss is introduced. Thus the improvement attributed to "CoT SFT" in the abstract and in Table 6 is not evidence for a benefit of intermediate reasoning; it is evidence primarily for the PLW loss re-weighting. A proper control is required, for example training a checkpoint to 10,000 steps without CoT data, or using a non-informative think section with the same PLW schedule, before any reasoning-specific benefit can be claimed. This concern is reinforced by the authors' own statement in §3.5 that "the intermediate reasoning steps did not demonstrate a direct causal influence on the final output," which contradicts the abstract's claim that CoT augmentation "foster[s] intermediate reflection."
  2. [§3.5, Table 2] The reward-function ablation is reported without repeated runs or any measure of variance. The differences among ckpt-10000 (13.42%), RF1+RF2 (13.15%), and RF1+RF2+RF5 (12.73%) are within about 0.7 percentage points on the development set, which may well be within run-to-run noise for RL training on a 1,500-hour corpus. Without multiple seeds or at least a significance test, the conclusion that RF5 contributes a meaningful self-correction gain is not established. The paper should report the number of runs and the standard deviation or confidence intervals for the key configurations, or explicitly state that these differences are not statistically tested.
minor comments (5)
  1. [§3.4, Eq. (2)] The loss formula and its weighting scheme are underspecified: the text says w_i = 1 for the <transcribe> section and w_i = PLW for the <think> section, but it does not state how structural tokens such as <LANG XX>, <transcribe>, and <think> themselves are weighted, nor whether the prompt tokens are fully masked. Clarifying the exact token weighting would make the method reproducible.
  2. [Table 1] The row label "ckpt-100000" appears to be a typo for "ckpt-10000" (ten zeros instead of four), which is confusing given the surrounding text.
  3. [Table 6] The caption uses "RLVF" while the text uses "RLVR" and "Dr. GRPO"; please standardize the acronym.
  4. [§4.2, Table 4] The decoding experiments include only a single run per configuration and do not report any confidence intervals; given that many configurations differ by well under one WER point, the non-monotonic beam-size trend may be noise. A brief statement acknowledging this would be appropriate.
  5. [References] Reference [6] is truncated mid-sentence ("serving over 90"), and reference [1] has an access date of 2024-03-20 while the challenge materials are dated 2025; please correct these bibliographic details.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical system paper with held-out evaluation; ablations are confounded but not circular.

full rationale

This is an empirical systems paper, not a derivation from assumptions. The final test-set numbers (11.57% Track 1, 17.67% Track 2) are measured against held-out challenge evaluation data, not computed from the training objective or from fitted parameters. The training pipeline uses ground-truth transcriptions as supervision and as verifiable reward signals, which is standard supervised/RL training, and the evaluation is independent of these labels. The development set is used for checkpoint selection and hyperparameter tuning (Section 2.3 and Tables 1-4), which is normal practice; it does not make the test-set result circular because the test set was not used for selection. The authors honestly report the negative ckpt-9000 result and the upper-bound context experiment (Section 4.1), and no load-bearing claim is justified solely by a self-citation chain: all cited prior work is external (SLAM-ASR, PLW, GRPO, etc.). The main weakness is that Table 1's comparison between ckpt-7000 and ckpt-10000 varies both CoT data, loss weighting, and step count, so the 'each component effective' claim is partially confounded; however, confounding is an experimental-design issue, not circularity. No step reduces by construction to its own input, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The central claims depend on a set of hand-picked tokens, loss weights, and hyperparameters that are not derived from theory and are validated only on the challenge dev set, with no code released. This makes the contributions engineering-specific rather than principled.

free parameters (5)
  • LoRA rank = 16
    Used in Stage 3 LoRA adaptation on both LLM and Whisper encoder; no sensitivity study reported.
  • PLW loss weight for <think> section = decayed 1.0 to 0.1 over 300 steps then held at 0.1
    Hand-chosen schedule in Stage 4; rescues CoT stage from 35.03% to 13.42% WER.
  • Reward weights in RLVR = RF2 weight 0.8, RF1/RF3/RF4/RF5 weight 0.5
    Weights chosen by experiment; final best combination is RF1+RF2+RF5.
  • VAD max segment time = 8 seconds
    Chosen by comparing 60s vs 8s; improves tcpWER by about 6.6 points.
  • RLVR sampling temperature = 0.5
    Controls diversity in rollout; fixed during experiments.
assumptions (5)
  • domain assumption Whisper large-v3-Turbo encoder output can be projected into Babel-9B-Chat embedding space with a 17.32M-parameter projector
    Section 2.1 architecture choice; no theoretical justification, validated only by final WER.
  • domain assumption Chain-of-thought data generated from Stage 3 model plus ground truth teaches self-correction
    Section 3.4; the assumption is that modeling 'hypothesis 1, error details, hypothesis 2' leads to better final transcription. This initially failed (35% WER) until PLW was added.
  • domain assumption Outcome-based rewards are more stable than process-based rewards for RLVR self-correction
    Section 3.5; supported by their experiments but stated as a conjecture, citing [27] and [28].
  • domain assumption The improvement from ckpt-7000 to ckpt-10000 is attributable to CoT and PLW rather than extra training steps
    Section 3.4 and Table 1; this is the load-bearing assumption behind the 'effectiveness of each component' claim, and it is confounded.
  • standard math Standard loss functions and WER/CER computations as evaluation metrics
    Equations 1 and 2; standard supervised learning assumptions.
invented entities (3)
  • <think> and </think> reasoning tokens
    purpose: Force the model to emit intermediate reflection before final transcription
    Only evaluated within this paper; no external benchmark isolates their contribution.
  • <LANG XX> language tokens
    purpose: Identify the language of each utterance
    Design choice for output structuring; effectiveness measured only inside the reported pipeline.
  • <speech> and <transcribe> delimiters
    purpose: Mark input speech boundaries and output transcription segments
    Engineering tokens; no independent evidence beyond in-paper ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seewo's Submission to MLC-SLM: Lessons learned from Speech Reasoning Language Models." pith.science (2026). https://pith.science/paper/VGTRBJPD

@misc{pith2026250613300,
  author       = {Pith},
  title        = {Pith review of: Seewo's Submission to MLC-SLM: Lessons learned from Speech Reasoning Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGTRBJPD}},
  note         = {Machine review of arXiv:2506.13300}
}
read the original abstract

This paper presents Seewo's systems for both tracks of the Multilingual Conversational Speech Language Model Challenge (MLC-SLM), addressing automatic speech recognition (ASR) and speaker diarization with ASR (SD-ASR). We introduce a multi-stage training pipeline that explicitly enhances reasoning and self-correction in speech language models for ASR. Our approach combines curriculum learning for progressive capability acquisition, Chain-of-Thought data augmentation to foster intermediate reflection, and Reinforcement Learning with Verifiable Rewards (RLVR) to further refine self-correction through reward-driven optimization. This approach achieves substantial improvements over the official challenge baselines. On the evaluation set, our best system attains a WER/CER of 11.57% for Track 1 and a tcpWER/tcpCER of 17.67% for Track 2. Comprehensive ablation studies demonstrate the effectiveness of each component under challenge constraints.

Figures

Figures reproduced from arXiv: 2506.13300 by the authors.

Figure 1
Figure 1. Model architecture referencing SLAM-ASR [2] 2. System overview This section details the core configuration of our system, in￾cluding model architecture, data processing, computational re￾sources and toolkits. All experiments used this configuration. 2.1. Foundation models Our speech language model architecture follows the SLAM￾ASR framework design [2]. SLAM-ASR provides a compre￾hensive evaluation of different encod… view at source ↗
Figure 2
Figure 2. Pipeline of speaker diarization 5. Speaker diarization pipeline Our best speaker diarization system uses the 3D-Speaker￾Toolkit [30], following the standard speaker diarization recipe without overlap detection. The system architecture comprises three key components: (1) a deep FSMN-based [31] voice activ￾ity detection (VAD) module for precise speech segment identi￾fication, (2) a ResNet-101 speaker embedding model f… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [1]

    The English subset contains approximately 500 hours of recordings from diverse regions, including British, American, Australian, In- dian, and Philippine English

    Introduction The MLC-SLM challenge focuses on multilingual conversa- tional speech recognition and speaker diarization tasks, encom- passing 11 languages: English (en), French (fr), German (de), Italian (it), Portuguese (pt), Spanish (es), Japanese (jp), Korean (ko), Russian (ru), Thai (th), and Vietnamese (vi). The English subset contains approximately 5...

    work page

  2. [2]

    Seewo's Submission to MLC-SLM: Lessons learned from Speech Reasoning Language Models

    System overview This section details the core configuration of our system, in- cluding model architecture, data processing, computational re- sources and toolkits. All experiments used this configuration. 2.1. Foundation models Our speech language model architecture follows the SLAM- ASR framework design [2]. SLAM-ASR provides a compre- hensive evaluation...

    work page Pith review arXiv 2025

  3. [3]

    Training pipeline of ASR system We adopt a multi-stage training pipeline where each stage builds upon the previous checkpoint to address specific challenges. While we present only the successful experiments here, our development process included numerous exploratory attempts that, despite not improving performance, provided valuable in- sights into the mo...

    work page 2000

  4. [4]

    Additional Experiments 4.1. Effect of Conversational Context on ASR Inspired by GEC-RAG [29], which demonstrated that con- textual information can enhance error correction in post- processing, we explored the impact of conversational context on ASR model performance in the MLC-SLM challenge. Given the conversational and long-form nature of the audio, we h...

    work page

  5. [5]

    Speaker diarization pipeline Our best speaker diarization system uses the 3D-Speaker- Toolkit [30], following the standard speaker diarization recipe without overlap detection. The system architecture comprises three key components: (1) a deep FSMN-based [31] voice activ- ity detection (V AD) module for precise speech segment identi- fication, (2) a ResNe...

    work page

  6. [6]

    Our approach achieves substantial improvements over the official challenge baselines

    Conclusions We presented a multi-stage training pipeline for multilingual conversational ASR, integrating LoRA, Chain-of-Thought data augmentation, and RLVR-based self-correction. Our approach achieves substantial improvements over the official challenge baselines. Table 6 and 7 summarize the results of our main submission to the evaluation set, demonstra...

    work page

  7. [7]

    Instruction fine-tuning: Does prompt loss matter?

    M. Huerta-Enochian and S. Y . Ko, “Instruction fine-tuning: Does prompt loss matter?” 2024

    work page 2024

  8. [8]

    Mlc-slm-baseline,

    M. Shen, “Mlc-slm-baseline,” https://github.com/mubingshen/ MLC-SLM-Baseline, 2025, accessed: 2024-03-20

    work page 2025

Show all 43 references
  1. [9]

    An embarrassingly simple ap- proach for llm with strong asr capacity,

    Z. Ma, G. Yang, Y . Yang, Z. Gao, J. Wang, Z. Du, F. Yu, Q. Chen, S. Zheng, S. Zhang, and X. Chen, “An embarrassingly simple ap- proach for llm with strong asr capacity,” 2024

    2024

  2. [10]

    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 Journal of Selecte...

    2022

  3. [11]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,” 2021

    2021

  4. [12]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” 2022

    2022

  5. [13]

    Y . Zhao, C. Liu, Y . Deng, J. Ying, M. Aljunied, Z. Li, L. Bing, H. P. Chan, Y . Rong, D. Zhao, and W. Zhang, “Babel: Open mul- tilingual large language models serving over 90

  6. [14]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” in2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 5220–5224

    2017

  7. [15]

    Understanding r1-zero-like training: A critical perspec- tive,

    Z. Liu, C. Chen, W. Li, P. Qi, T. Pang, C. Du, W. S. Lee, and M. Lin, “Understanding r1-zero-like training: A critical perspec- tive,” 2025

    2025

  8. [16]

    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 . K. Li, Y . Wu, and D. Guo, “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” 2024

    2024

  9. [17]

    Cn-celeb: multi-genre speaker recogni- tion,

    L. Li, R. Liu, J. Kang, Y . Fan, H. Cui, Y . Cai, R. Vipperla, T. F. Zheng, and D. Wang, “Cn-celeb: multi-genre speaker recogni- tion,” 2020

    2020

  10. [18]

    V oxblink: A large scale speaker verification dataset on camera,

    Y . Lin, X. Qin, G. Zhao, M. Cheng, N. Jiang, H. Wu, and M. Li, “V oxblink: A large scale speaker verification dataset on camera,” 2023

    2023

  11. [19]

    V oxblink2: A 100k+ speaker recognition corpus and the open- set speaker-identification benchmark,

    Y . Lin, M. Cheng, F. Zhang, Y . Gao, S. Zhang, and M. Li, “V oxblink2: A 100k+ speaker recognition corpus and the open- set speaker-identification benchmark,” 2024

    2024

  12. [20]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  13. [21]

    Qwen2.5 technical report,

    Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, ...

    2025

  14. [22]

    MUSAN: A Music, Speech, and Noise Corpus,

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A Music, Speech, and Noise Corpus,” 2015, arXiv:1510.08484v1

    2015 arXiv

  15. [23]

    We speech transcript, llm based speech recog- nition/transcript in 300 lines of code,

    W. Team, “We speech transcript, llm based speech recog- nition/transcript in 300 lines of code,” https://github.com/ wenet-e2e/west, 2024, accessed: 2024-03-20

    2024

  16. [24]

    Transform- ers: State-of-the-art natural language processing,

    T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y . Jernite, J. Plu, C. Xu, T. L. Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush, “Transform- ers: State-of-the-art n...

    2020

  17. [25]

    Accelerate: Training and in- ference at scale made simple, efficient and adaptable

    S. Gugger, L. Debut, T. Wolf, P. Schmid, Z. Mueller, S. Man- grulkar, M. Sun, and B. Bossan, “Accelerate: Training and in- ference at scale made simple, efficient and adaptable.” https: //github.com/huggingface/accelerate, 2022

    2022

  18. [26]

    Peft: State-of-the-art parameter-efficient fine-tuning methods,

    S. Mangrulkar, S. Gugger, L. Debut, Y . Belkada, S. Paul, and B. Bossan, “Peft: State-of-the-art parameter-efficient fine-tuning methods,” https://github.com/huggingface/peft, 2022

    2022

  19. [27]

    Trl: Trans- former reinforcement learning,

    L. von Werra, Y . Belkada, L. Tunstall, E. Beeching, T. Thrush, N. Lambert, S. Huang, K. Rasul, and Q. Gallou´edec, “Trl: Trans- former reinforcement learning,” https://github.com/huggingface/ trl, 2020

    2020

  20. [28]

    Reasoning language models: A blueprint,

    M. Besta, J. Barth, E. Schreiber, A. Kubicek, A. Catarino, R. Ger- stenberger, P. Nyczyk, P. Iff, Y . Li, S. Houliston, T. Sternal, M. Copik, G. Kwa ´sniewski, J. M ¨uller, Łukasz Flis, H. Eberhard, Z. Chen, H. Niewiadomski, and T. Hoefler, “Reasoning language models: A bluepr...

    2025

  21. [29]

    Unified language-vision pretraining in llm with dynamic discrete visual tokenization,

    Y . Jin, K. Xu, K. Xu, L. Chen, C. Liao, J. Tan, Q. Huang, B. Chen, C. Lei, A. Liu, C. Song, X. Lei, D. Zhang, W. Ou, K. Gai, and Y . Mu, “Unified language-vision pretraining in llm with dynamic discrete visual tokenization,” 2024

    2024

  22. [30]

    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,” 2021

    2021

  23. [31]

    S 2r: Teaching llms to self-verify and self- correct via reinforcement learning,

    R. Ma, P. Wang, C. Liu, X. Liu, J. Chen, B. Zhang, X. Zhou, N. Du, and J. Li, “S 2r: Teaching llms to self-verify and self- correct via reinforcement learning,” 2025

    2025

  24. [32]

    Training language mod- els to self-correct via reinforcement learning,

    A. Kumar, V . Zhuang, R. Agarwal, Y . Su, J. D. Co-Reyes, A. Singh, K. Baumli, S. Iqbal, C. Bishop, R. Roelofs, L. M. Zhang, K. McKinney, D. Shrivastava, C. Paduraru, G. Tucker, D. Precup, F. Behbahani, and A. Faust, “Training language mod- els to self-correct via reinforcemen...

    2024

  25. [33]

    Instruction tuning with loss over instructions,

    Z. Shi, A. X. Yang, B. Wu, L. Aitchison, E. Yilmaz, and A. Lipani, “Instruction tuning with loss over instructions,” 2024

    2024

  26. [34]

    Glore: When, where, and how to improve llm reasoning via global and local refinements,

    A. Havrilla, S. Raparthy, C. Nalmpantis, J. Dwivedi-Yu, M. Zhu- ravinskyi, E. Hambro, and R. Raileanu, “Glore: When, where, and how to improve llm reasoning via global and local refinements,” 2024

    2024

  27. [35]

    The ResNet-101 model was specifically trained on a diverse corpus comprising CnCeleb1, CnCeleb2 [10], V oxBlink [11], and V oxBlink2 [12] datasets

    from ModelScope [36]. The ResNet-101 model was specifically trained on a diverse corpus comprising CnCeleb1, CnCeleb2 [10], V oxBlink [11], and V oxBlink2 [12] datasets. A critical optimization involved adjusting the V AD mod- ule’smax single segment timeparameter. The default...

  28. [36]

    Gec-rag: Improving generative error correction via retrieval-augmented generation for automatic speech recognition systems,

    A. Robatian, M. Hajipour, M. R. Peyghan, F. Rajabi, S. Amini, S. Ghaemmaghami, and I. Gholampour, “Gec-rag: Improving generative error correction via retrieval-augmented generation for automatic speech recognition systems,” 2025

    2025

  29. [37]

    3d-speaker-toolkit: An open source toolkit for multi-modal speaker verification and diarization,

    Y . Chen, S. Zheng, H. Wang, L. Chenget al., “3d-speaker-toolkit: An open source toolkit for multi-modal speaker verification and diarization,” 2025

    2025

  30. [38]

    Deep-fsmn for large vo- cabulary continuous speech recognition,

    S. Zhang, M. Lei, Z. Yan, and L. Dai, “Deep-fsmn for large vo- cabulary continuous speech recognition,” in2018 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5869–5873

    2018

  31. [39]

    Advancing speaker embedding learning: Wespeaker toolkit for research and produc- tion,

    S. Wang, Z. Chen, B. Han, H. Wang, C. Liang, B. Zhang, X. Xi- ang, W. Ding, J. Rohdin, A. Silnovaet al., “Advancing speaker embedding learning: Wespeaker toolkit for research and produc- tion,”Speech Communication, vol. 162, p. 103104, 2024

    2024

  32. [40]

    A spectral clustering approach to speaker diarization

    H. Ning, M. Liu, H. Tang, and T. S. Huang, “A spectral clustering approach to speaker diarization.” inInterspeech. Citeseer, 2006

    2006

  33. [41]

    An enhanced res2net with local and global feature fusion for speaker verification,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, Q. Chen, and J. Qi, “An enhanced res2net with local and global feature fusion for speaker verification,” 2023

    2023

  34. [42]

    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,”arXiv preprint arXiv:2303.00332

    arXiv

  35. [43]

    Modelscope: bring the notion of model-as-a-service to life

    T. M. Team, “Modelscope: bring the notion of model-as-a-service to life.” https://github.com/modelscope/modelscope, 2023

    2023

Pith tools

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