Pith. sign in

REVIEW 2 major objections 5 minor 29 references

Language-Specialized Multi-Teacher On-Policy Distillation for Multilingual LLM-Based ASR

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A multilingual ASR student distilled from language-specialized RL teachers can surpass every teacher it learned from.

desk verdict Worth a read, but the headline 'beyond the teacher' claim has a load-bearing confound: the 20k-step distilled student is compared against 6k-step teachers. read the letter →

arxiv 2608.03610 v1 pith:PAD6ICP3 submitted 2026-08-04 cs.CL cs.SDeess.AS

Yuan Xie , Jiaqi Song , Xianliang Wang , Ming Lei , Jie Gao , Jie Wu This is my paper
classification cs.CLcs.SDeess.AS
keywords multilingualASRLLM-basedon-policydistillationmulti-teacherlanguageroutingreinforcementlearningwithverifiablerewardsacousticprefix
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

This paper sets out to show that a single multilingual ASR model can recognize speech better than any of the teachers that trained it, if the teachers are specialized per language and their supervision is distilled on the student's own outputs. The proposed method, LS-MOPD, first trains separate reinforcement-learning teachers for Mandarin, Chinese dialects, and English, then routes each utterance to the best-ranked teachers and applies weighted token-level distillation into one 2.3B generalist student. On seven test sets, that student, trained on only 50k post-training utterances, beats every individual teacher and also beats larger open-source ASR systems trained on far more speech. The paper further shows that a static, shared acoustic encoder distills better than letting each teacher use its own encoder, and that lower-ranked teachers still contribute useful complementary supervision.

What carries the argument

The load-bearing machinery is language-routed multi-teacher on-policy distillation with DAPO-trained specialists. DAPO is a token-level reinforcement-learning algorithm with group-normalized advantages and dynamic sampling; it converts an ASR-specific accuracy reward into language-specialized teachers. During distillation, the student samples its own transcriptions, each utterance is routed by its language label to the top-K ranked teachers for that category, and the student is trained with weighted token-level reverse KL against those teachers. The input-language category determines which teachers supervise, while the acoustic-prefix configuration—static shared frozen encoder versus dynamic

What would settle it

Take the same 50k-utterance pool and deliberately corrupt an increasing fraction of the routing labels, or feed code-switched utterances through an automatic language tagger: if LS-MOPD's advantage over a single generalist teacher survives heavy label corruption, language routing is not the causal mechanism; if performance collapses, the routing assumption is load-bearing.

Watch

Extended reading notes

Core claim

The central discovery is that a student can exceed the empirical performance envelope of all its supervising teachers in multilingual ASR, not merely interpolate between them. LS-MOPD trains language-specialized teachers with a token-level reinforcement-learning algorithm and an ASR-specific accuracy reward, ranks those teachers separately for each language category, and distills the top three into the student with a weighted reverse-KL loss on student-generated rollouts. Across offline and streaming benchmarks, the student's average CER/WER falls below the best-teacher oracle on almost every subset, and the analysis identifies a measurable share of tokens that the student transcribes correc

Load-bearing premise

The load-bearing premise is that every training utterance is correctly labeled into one of three language groups before distillation and that those groups are enough to pick a good supervising teacher; if labels are noisy or speech mixes languages, the student is taught by teachers not suited to the input.

Editorial extensions

If this is right

  • If the result holds up, a small generalist student post-trained on a 50k-utterance pool can match or beat much larger ASR systems on Mandarin, Cantonese, and English, making distillation an economical alternative to scaling training data.
  • The static-prefix result implies that for speech-conditioned LLMs, keeping teacher and student acoustic representations consistent can matter more than peak teacher quality.
  • Weighted multi-teacher supervision acts as a regularizer, with the largest gains in streaming decoding where uncertainty is higher, so lower-ranked teachers are worth retaining despite weaker standalone performance.
  • OPD–CE gradient alignment is a workable, low-cost predictor of how much a distillation configuration will improve error rates, useful for tuning teacher sets without full evaluations.
  • The existence of tokens the student gets right while all teachers are wrong indicates that distillation can create capabilities not present in any individual supervisor.

Reading between the lines

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

  • Editorial inference: The method assumes a correct language label for every utterance; a natural stress test is to corrupt a fraction of routing labels or use code-switched utterances and watch whether the advantage over a single generalist teacher shrinks.
  • Editorial inference: The same recipe—specialized teachers plus consistency-preserving on-policy distillation—could apply to other LLM tasks with continuous conditioning inputs, such as image- or video-conditioned generation, where a frozen conditioning encoder plays the static-prefix role.
  • Editorial inference: Because the evaluation covers one backbone and three language groups, an open question is whether the beyond-all-teachers effect grows with the number and heterogeneity of language-specialized teachers.
  • Editorial inference: The 'student-only correct' tokens could be studied by error class (homophones, prosody, rare words) to see whether the student is genuinely combining teacher knowledge or merely benefiting from test-set regularities.
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. The paper proposes LS-MOPD, a post-training framework for multilingual LLM-based ASR. It first trains three language-specialized DAPO-based RL teachers (Mandarin, Chinese dialects, English) and a generalist teacher on a 50k-utterance pool. It then distills these teachers into a single 2.3B student using token-level reverse-KL on student-generated rollouts, routing each example to the top-K teachers for its language category. Two acoustic-prefix configurations (static and dynamic) are compared, and offline/streaming results are reported on WenetSpeech, KeSpeech, WenetSpeech-Yue, and LibriSpeech. The central empirical claim is that the distilled student outperforms every individual RL teacher and even the per-benchmark best-teacher oracle, and also outperforms several larger or more heavily trained open-source models on average.

Significance. If the claims hold, this is a useful and timely contribution: it is the first application of multi-teacher on-policy distillation to ASR, and the static/dynamic acoustic-prefix analysis with gradient-alignment diagnostics offers a plausible mechanistic story for when distillation works. The paper also makes a good-faith effort to compare against multiple public baselines and reports bootstrap uncertainties in Fig. 2. However, the headline 'beyond all teachers' claim is currently threatened by an unequal training budget between teachers and student, and by the absence of uncertainty quantification in the main tables. An equal-step control and variance reporting would substantially strengthen the paper.

major comments (2)
  1. [Sec. 4.2 and Tables 1-2] The central comparison is not controlled for optimization budget. RL teachers are trained for at most 6k DAPO steps (Sec. 4.2), while the MOPD student trains for 20k steps, and Fig. 2(a) shows teacher reward still rising at 6k. There is no DAPO-trained generalist run for 20k steps and no SFT/CE baseline on the same 50k pool at matched steps. Therefore the reported gains over 'RL (Best-Teacher Oracle)' and the RL baseline may reflect additional gradient steps on the same data rather than language-routed multi-teacher supervision. Please add an equal-step (or equal-FLOP) RL generalist, or extend the best teacher to 20k steps, and report per-configuration compute. Without this, the claim that LS-MOPD 'surpasses the empirical performance envelope' is not attributable to the proposed mechanism.
  2. [Tables 1-2] The main results are reported as single point estimates with no seed variance, confidence intervals, or significance tests. The claimed 'consistent' improvement includes many margins of only 0.02 on the average (e.g., Table 2, dynamic streaming, LS-MOPD 2-teacher: 5.36 vs. oracle 5.36) and several per-benchmark ties. Since the core message is that the student surpasses all teachers, the evaluation needs multiple seeds or bootstrap CIs to show that these differences are not noise. Fig. 2 already uses bootstrap resampling; the same rigor should be applied to Tables 1 and 2.
minor comments (5)
  1. [Sec. 3.2] It is not stated whether the language category ℓ(x) comes from dataset metadata or from an automatic language classifier. Since code-switched or heavily dialectal speech may violate the routing assumption, this should be clarified.
  2. [Sec. 4.2 / Hyperparameters] Several hyperparameters (K=3, λ weights, reward exponent −2, group size G=8, sampling ratios) are introduced without sensitivity analysis. At least one ablation varying λ or K would help the reader understand robustness.
  3. [Fig. 2] The caption says error bars are estimated from 10k bootstrap resamples, but the resampling unit (utterance, token, or trajectory) is not specified. Please also explain why the same uncertainty estimation is not applied to Tables 1 and 2.
  4. [Tables 1-2] Minor formatting: 'WenetSpeech-Yue' is sometimes written as 'WeNet-Yue', and the column alignment is hard to read. Please standardize.
  5. [Conclusion] The limitations paragraph acknowledges that the dynamic acoustic-prefix configuration is preliminary. This is a good practice; consider reflecting it earlier in the paper so the abstract is not overclaimed relative to the current evidence.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the student objective is a weighted reverse-KL to independently trained teachers, and the 'beyond-teacher' result is empirical; only minor non-load-bearing self-citation of the authors' NIM4-ASR backbone/reward appears.

full rationale

The paper's derivation chain is self-contained. The student is trained by minimizing a weighted sum of reverse-KL losses (Eqs. 3–4) against teacher distributions, with teachers independently trained by DAPO RL (Eqs. 1–2) and ranked on held-out validation sets. The central claim—that LS-MOPD surpasses the best-teacher oracle—is an empirical result on official test sets, not a quantity defined in terms of the training objective. No fitted parameter is relabeled as a prediction: the aggregation weights λ_k are fixed a priori ((1,0,0), (0.6,0.4,0), (0.6,0.2,0.2)), and top-K routing uses validation, not test, performance. The only self-citation is the use of the authors' NIM4-ASR backbone and its accuracy reward (Eq. 1, [10]); this is an input to the method rather than a validation of the central claim, so it is not load-bearing. The 'beyond all teachers' observation is not circular by construction: the student is initialized from the base model and trained with imperfect optimization, so it can assign probability mass differently from any single teacher; Fig. 3(b) directly measures positions where the student is correct while all teachers are wrong. The skeptics' concerns—unequal training steps (≤6k for RL teachers vs. 20k for MOPD), absence of an equal-compute RL/SFT control, and reliance on known language labels ℓ(x)—are correctness/confound risks rather than circularity, because they do not make the output equal to the input by definition. The paper's acknowledged limitations (single backbone, narrow language coverage, preliminary dynamic-prefix design) are external-validity issues, not circular steps. Overall, no equation-level or definitional circularity; score 2 reflects only minor non-load-bearing self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. The load-bearing choices are the hand-set aggregation weights, the 20% data-mixing ratio, and the assumptions that language labels are known and that reverse-KL on-policy distillation transfers across acoustic-prefix conditions. These are design choices and modeling assumptions, not fitted constants in a derivation.

free parameters (5)
  • Multi-teacher loss weights lambda_k = (0.6, 0.2, 0.2) for top-3; (0.6, 0.4) for top-2; (1, 0, 0) for top-1
    Chosen by hand, not searched. Eq. (4). The weights directly control how much each teacher influences the student, so the central aggregation result depends on this choice.
  • Cross-category sampling ratio for specialist teachers = 20%
    Each specialist teacher trains on 100% of its target category plus a randomly sampled 20% of each other category. This ratio determines teacher specialization; ad hoc.
  • Top-K number of supervising teachers = K=3 default
    Set by default; the paper compares 1, 2, and 3 teachers. The claim that lower-ranked teachers help rests on this choice.
  • Reward exponent coefficient (Eq. 1) = -2
    The accuracy reward is exp(-2 * error rate), adopted from prior work [10]. A fixed scaling constant; not tuned here, but it shapes the RL teachers' behavior.
  • Rollout group size G for RL = G=8
    Number of hypotheses sampled per utterance for advantages; standard choice but affects teacher quality. Not the paper's main focus.
assumptions (4)
  • domain assumption On-policy distillation with a reverse-KL objective is a valid training signal that lets a student improve beyond its teacher's error envelope.
    Invoked in Eq. (3) and throughout Sec. 3.2; the entire 'emergent gains' claim depends on this being true for ASR.
  • domain assumption The input language category ℓ(x) is known and correct at distillation time, and coarse routing into Mandarin, Chinese dialects, and English is sufficient to choose useful teachers.
    Sec. 3.2 uses ℓ(x) to select the supervising teacher. No language identification or code-switching robustness is discussed; wrong labels would feed mismatched supervision.
  • domain assumption Validation-based teacher rankings generalize to the official test sets.
    Sec. 3.2 ranks teachers on held-out validation to define top-K routing; if the ranking flips on test, the student would be trained toward teachers that are not the best for the benchmark.
  • standard math Standard reinforcement learning and KL divergence identities (DAPO objective, Eq. 2) hold.
    Adopted from cited prior work [8]. No new math; standard policy-gradient machinery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Language-Specialized Multi-Teacher On-Policy Distillation for Multilingual LLM-Based ASR." pith.science (2026). https://pith.science/paper/PAD6ICP3

@misc{pith2026260803610,
  author       = {Pith},
  title        = {Pith review of: Language-Specialized Multi-Teacher On-Policy Distillation for Multilingual LLM-Based ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PAD6ICP3}},
  note         = {Machine review of arXiv:2608.03610}
}
read the original abstract

Modern LLM-based ASR systems have established multilingual capability as a standard feature, leveraging large-scale multilingual corpora and LLMs' cross-lingual knowledge to achieve competitive performance across multilingual benchmarks. However, joint modeling of languages with heterogeneous acoustic, phonological, and lexical characteristics inevitably introduces optimization conflicts, undermining language-wise specialization. To address this challenge, we propose Language-Specialized Multi-Teacher On-Policy Distillation (LS-MOPD), which decouples language-specific knowledge acquisition from multilingual capability integration: language-specialized teachers are independently optimized via reinforcement learning (RL), after which their expertise is integrated into a generalist multilingual student through language routing and token-level multi-teacher distillation, thereby reducing direct cross-lingual optimization conflicts. We further explore two acoustic-prefix configurations, static and dynamic, to examine how teacher--student prefix consistency influences the efficacy of on-policy distillation. Experiments on benchmarks covering Mandarin, Mandarin subdialects, Cantonese, and English demonstrate that LS-MOPD substantially outperforms RL baselines and consistently surpasses the empirical performance envelope defined by best-performing RL teachers, revealing its potential to generalize beyond all teachers in multilingual ASR.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages

  1. [1]

    INTRODUCTION LLM-based ASR has rapidly advanced as the mainstream paradigm in speech recognition, capitalizing on the extensive model capacity and rich linguistic priors of pretrained LLMs to achieve impressive performance across dozens of lan- guages and dialects. Frontier systems such as Seed-ASR [1], Fun-ASR [2], and Qwen3-ASR [3] primarily rely on sca...

    work page Pith review arXiv 2026

  2. [2]

    RELATED WORK Multilingual ASR.Extending multilingual capability has become a major focus of modern ASR systems, with exist- ing efforts following two main directions. The first lever- ages industrial-scale data expansion: systems such as Seed- ASR [1], Fun-ASR [2], Qwen3-ASR [3], FireRedASR2S [9], and NIM4-ASR [10] achieve strong multilingual perfor- manc...

    work page

  3. [3]

    T eacher 1 supervision T eacher Model m ( Encoder + adaptor ) T eacher Model m (LLM decoder) T oken 1 T oken 2 T oken n

    Optimization T op-1 routing T eacher Model 1 ( Encoder + adaptor ) T eacher Model 1 (LLM decoder) T oken 1 T oken 2 T oken n ...... T eacher 1 supervision T eacher Model m ( Encoder + adaptor ) T eacher Model m (LLM decoder) T oken 1 T oken 2 T oken n ...... T eacher m supervision

    work page

  4. [4]

    Optimization embeddingsprompt embeddingsprompt Optimization only when using dynamic acoustic pr efix Reverse KL Reverse KL W eighted Sum (b) LS-MOPD with Dynamic / Static acoustic pr efix ... ... (active only with a dynamic acoustic pr efix) Dynamic acoustic prefix ... static mode dynamic mode unified Input speech Static acoustic pr efix : r euse student'...

    work page

  5. [5]

    Advanced Open-source Models

    METHODS 3.1. Language-Specialized RL Teacher Training Prior to conducting MOPD, we construct a suite of teacher models specialized for distinct language categories. Starting from a base modelπ θ0 built upon the encoder–adaptor–LLM architecture, we perform reinforcement learning with verifi- able rewards (RLVR) based on DAPO [8]. For a given utter- ancexwi...

    work page

  6. [6]

    RL (Best- Teacher Oracle)

    EXPERIMENTS 4.1. Experimental Setup The base student model in this work is built upon an in- house backbone pretrained on approximately 560k hours of labeled speech, comprising a 0.6B FireRed Conformer [9], a 4×downsampling linear adaptor, and a Qwen3-1.7B LLM decoder. For RL and MOPD, we construct a 50k-utterance multilingual training pool by sampling fr...

    work page

  7. [7]

    CONCLUSION This work proposes a multi-teacher on-policy distillation framework for multilingual LLM-based ASR. We systemat- ically investigate the impact of static and dynamic acoustic- prefix configurations and integrate language-specialized teachers through language-routed, weighted supervision. Ex- perimental results across both offline and streaming s...

    work page

  8. [8]

    Seed-asr: Understanding di- verse speech and contexts with llm-based speech recog- nition,

    Ye Bai, Jingping Chen, Jitong Chen, Wei Chen, Zhuo Chen, Chuang Ding, Linhao Dong, Qianqian Dong, Yu- jiao Du, Kepan Gao, et al., “Seed-asr: Understanding di- verse speech and contexts with llm-based speech recog- nition,”arXiv preprint arXiv:2407.04675, 2024

    arXiv 2024

Show all 29 references
  1. [9]

    Fun-asr technical report,

    Keyu An, Yanni Chen, Zhigao Chen, Chong Deng, Zhi- hao Du, Changfeng Gao, Zhifu Gao, Bo Gong, Xian- gang Li, Yabin Li, et al., “Fun-asr technical report,” arXiv preprint arXiv:2509.12508, 2025

    2025

  2. [10]

    Qwen3-asr technical re- port,

    Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, et al., “Qwen3-asr technical re- port,”arXiv preprint arXiv:2601.21337, 2026

    2026 arXiv

  3. [11]

    Mimo-v2-flash technical report,

    Bangjun Xiao, Bingquan Xia, Bo Yang, Bofei Gao, Bowen Shen, Chen Zhang, Chenhong He, Chiheng Lou, Fuli Luo, Gang Wang, et al., “Mimo-v2-flash technical report,”arXiv preprint arXiv:2601.02780, 2026

    2026 arXiv

  4. [12]

    Mopd: Multi-teacher on-policy distillation for ca- pability integration in llm post-training,

    Wenhan Ma, Jianyu Wei, Liang Zhao, Hailin Zhang, Bangjun Xiao, Lei Li, Qibin Yang, Bofei Gao, Yudong Wang, Rang Li, Jinhao Dong, Zhifang Sui, and Fuli Luo, “Mopd: Multi-teacher on-policy distillation for ca- pability integration in llm post-training,”arXiv preprint arXiv:2606....

    2026 arXiv

  5. [13]

    Nemotron- cascade 2: Post-training llms with cascade rl and multi-domain on-policy distillation,

    Zhuolin Yang, Zihan Liu, Yang Chen, Wenliang Dai, Boxin Wang, Sheng-Chieh Lin, Chankyu Lee, Yangyi Chen, Dongfu Jiang, Jiafan He, et al., “Nemotron- cascade 2: Post-training llms with cascade rl and multi-domain on-policy distillation,”arXiv preprint arXiv:2603.19220, 2026

    2026

  6. [14]

    Deepseek- v4: Towards highly efficient million-token context in- telligence,

    Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al., “Deepseek- v4: Towards highly efficient million-token context in- telligence,”arXiv preprint arXiv:2606.19348, 2026

    2026

  7. [15]

    Dapo: An open-source llm reinforcement learning system at scale,

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xi- aochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gao- hong Liu, Lingjun Liu, et al., “Dapo: An open-source llm reinforcement learning system at scale,”Advances in Neural Information Processing Systems, vol. 38, pp. 113222–113244, 2025

    2025

  8. [16]

    Fireredasr2s: A state-of-the-art industrial-grade all-in-one automatic speech recognition system,

    Kaituo Xu, Yan Jia, Kai Huang, Junjie Chen, Wen- peng Li, Kun Liu, Feng-Long Xie, Xu Tang, and Yao Hu, “Fireredasr2s: A state-of-the-art industrial-grade all-in-one automatic speech recognition system,”arXiv preprint arXiv:2603.10420, 2026

    2026

  9. [17]

    Nim4-asr: Towards efficient, robust, and customizable real-time llm-based asr,

    Yuan Xie, Jiaqi Song, Guang Qiu, Xianliang Wang, Kai Qiao, Junfeng Yuan, Shengqing Liu, Yi Zhang, Bowen Chen, Ming Lei, et al., “Nim4-asr: Towards efficient, robust, and customizable real-time llm-based asr,”arXiv preprint arXiv:2604.18105, 2026

    2026 arXiv

  10. [18]

    Mosa: Mixtures of simple adapters outperform monolithic approaches in llm-based multilingual asr,

    Junjie Li, Jing Peng, Yangui Fang, Shuai Wang, and Kai Yu, “Mosa: Mixtures of simple adapters outperform monolithic approaches in llm-based multilingual asr,” in2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 18822–18826

    2026

  11. [19]

    Enhancing multilingual llm-based asr with mixture of experts and dynamic downsampling,

    Guodong Lin, Ziqi Chen, Yuxiang Fu, Ke Li, and Wei- Qiang Zhang, “Enhancing multilingual llm-based asr with mixture of experts and dynamic downsampling,” in2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 18807–18811

    2026

  12. [20]

    Language-aware distillation for multilingual instruction-following speech llms with asr- only supervision,

    Shreyas Gopal, Donghang Wu, Ashutosh Anshul, Yeo Yue Heng, Yizhou Peng, Haoyang Li, Hexin Liu, and Eng Siong Chng, “Language-aware distillation for multilingual instruction-following speech llms with asr- only supervision,”arXiv preprint arXiv:2603.07025, 2026

    2026 arXiv

  13. [21]

    On-policy distillation of language models: Learning from self-generated mistakes,

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Pi- otr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem, “On-policy distillation of language models: Learning from self-generated mistakes,” in International Conference on Learning Representations, 2024

    2024

  14. [22]

    Step-audio 2 technical report,

    Boyong Wu, Chao Yan, Chen Hu, Cheng Yi, Chengli Feng, Fei Tian, Feiyu Shen, Gang Yu, Haoyang Zhang, Jingbei Li, et al., “Step-audio 2 technical report,”arXiv preprint arXiv:2507.16632, 2025

    2025 arXiv

  15. [23]

    Qwen3-omni technical report,

    Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, et al., “Qwen3-omni technical report,”arXiv preprint arXiv:2509.17765, 2025

    2025 arXiv

  16. [24]

    Mimo-v2.5-asr: Robust speech recognition across languages, dialects, and com- plex acoustic scenarios,

    LLM-Core-Team Xiaomi, “Mimo-v2.5-asr: Robust speech recognition across languages, dialects, and com- plex acoustic scenarios,” 2026

    2026

  17. [25]

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

    Binbin Zhang, Hang Lv, Pengcheng Guo, Qijie Shao, Chao Yang, Lei Xie, Xin Xu, Hui Bu, Xiaoyu Chen, Chenchen Zeng, et al., “Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” in2022 IEEE International Conference on Acoustics, Speech and Signal Pro...

    2022

  18. [26]

    Kespeech: An open source speech dataset of mandarin and its eight subdialects,

    Zhiyuan Tang, Dong Wang, Yanguang Xu, Jianwei Sun, Xiaoning Lei, Shuaijiang Zhao, Cheng Wen, Xingjun Tan, Chuandong Xie, Shuran Zhou, et al., “Kespeech: An open source speech dataset of mandarin and its eight subdialects,” inThirty-fifth conference on neural in- formation proc...

    2021

  19. [27]

    Wenetspeech- yue: A large-scale cantonese speech corpus with multi- dimensional annotation,

    Longhao Li, Zhao Guo, Hongjie Chen, Yuhang Dai, Ziyu Zhang, Hongfei Xue, Tianlun Zuo, Chengyou Wang, Shuiyuan Wang, Xin Xu, et al., “Wenetspeech- yue: A large-scale cantonese speech corpus with multi- dimensional annotation,” inProceedings of the AAAI Conference on Artificial ...

    2026

  20. [28]

    Librispeech: an asr corpus based on public domain audio books,

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and San- jeev Khudanpur, “Librispeech: 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

  21. [29]

    Zero: Memory optimizations toward training trillion parameter models,

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He, “Zero: Memory optimizations toward training trillion parameter models,” inSC20: interna- tional conference for high performance computing, net- working, storage and analysis. IEEE, 2020, pp. 1–16

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.