Pith. sign in

REVIEW 4 major objections 4 minor 25 references

On the MLC-SLM 2026 two-speaker task, a three-stage adaptation recipe reduces Qwen3-ASR-1.7B's average tcpMER from 30.53 on dev to 23.70, and to 17.97 on eval.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 07:54 UTC pith:R5ZPPT66

load-bearing objection Useful challenge-system report whose headline number is internally consistent but measured on a local scoring pipeline; the adaptation gain likely holds, but the abstract overstates it as an official score. the 4 major comments →

arxiv 2607.08208 v2 pith:R5ZPPT66 submitted 2026-07-09 cs.CL

Diarization-Guided Qwen-ASR Adaptation for Multilingual Two-Speaker Conversational Speech

Hao Wu , RongQi Han , Zhen Wang , Wei Liang , Wei Xu This is my paper
classification cs.CL
keywords multilingual speech recognitionspeaker diarizationspeech language modelLoRATTS synthetic speech augmentationGRPO reinforcement learningtcpMERtwo-speaker conversational speech
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a three-stage adaptation recipe—full supervised fine-tuning, LoRA on synthetic TTS speech, and GRPO reinforcement learning—brings Qwen3-ASR-1.7B from 30.53 to 23.70 average tcpMER on the MLC-SLM 2026 Task 1 development set, and to 17.97 on the official evaluation set, for multilingual two-speaker conversational speech across 21 language/region conditions. The development-set gain comes from the authors' local scoring pipeline; the official judge's eval number is the externally validated result. If the claim holds, it shows a 1.7B speech language model can be made competitive for code-switching and low-resource languages like Urdu and Tagalog without architectural changes, and that supervised fine-tuning, not synthetic data or RL, provides the bulk of the improvement.

Core claim

The paper's central claim is that Qwen3-ASR-1.7B, at 30.53 average tcpMER on the MLC-SLM 2026 dev set under a shared diarization output, can be adapted by a three-stage recipe to 23.70: full supervised fine-tuning on official data, LoRA on 234,333 synthetic TTS items, and GRPO reinforcement learning with a rule-based reward. The ablation isolates the contributions: full SFT delivers 6.23 of the 6.83-point gain, synthetic-speech LoRA adds 0.52 (Urdu 102.84→19.77, Tagalog 49.49→33.90), and RL adds 0.08 by stabilizing outputs. The authors read this as SFT aligning the model to the challenge's two-speaker conversational style, synthetic speech broadening hard-language coverage, and RL acting as

What carries the argument

The central machinery is the three-stage adaptation recipe applied to the Qwen3-ASR-1.7B speech language model: (1) full supervised fine-tuning (SFT) on the official training set, (2) low-rank adaptation (LoRA) on synthetic speech generated by a three-pipeline TTS framework, and (3) Group Relative Policy Optimization (GRPO), a critic-free reinforcement-learning update, with a rule-based reward combining word/character error rate with penalties for hallucination, repetition, and length deviation. The diarization front end—FSMN-VAD, CAMPPlus speaker embeddings, and spectral clustering fixed to two speakers, followed by RTTM-based cutting—turns the recording into speaker-attributed segments, bu

Load-bearing premise

The load-bearing premise is that the authors' local implementation of tcpMER, which they use for the development set and explicitly label 'our local scoring pipeline' (§1, §4.1), reproduces the official scoring protocol; if the normalization, permutation search, or per-language averaging differs, every dev-set number changes, leaving the official eval result (17.97) as the only fully validated figure.

What would settle it

Run the submitted system through the official MLC-SLM 2026 judge on the development set and compare the returned tcpMER with the reported 23.70; agreement would validate the local pipeline, while a material discrepancy would invalidate the dev-set results and the ablation decomposition.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The adapted system beats both official ASR baselines under the same diarization-aware protocol: Whisper-large-v3 averages 29.94 tcpMER and Omniasr-LLM-7B-v2 averages 31.52, versus the submitted system's 23.70 on dev and 17.97 on eval.
  • Full SFT accounts for 6.23 of the 6.83 total dev-set gain, so for challenge-style adaptation of speech LLMs, high-quality supervised data is the dominant lever; synthetic LoRA and RL are refinements.
  • The largest per-language gains are in the weakest languages—Urdu 102.84→19.77, Tagalog 49.49→33.90—indicating the recipe specifically repairs under-represented acoustic-linguistic confusions.
  • GRPO with rule-based rewards (WER/CER plus hallucination, repetition, and length penalties) suppresses pathological outputs on long, noisy segments, offering a stable, human-feedback-free RL recipe for ASR.
  • Because the same diarization output is shared across all ablation runs, the measured improvement is attributable to recognizer adaptation alone, not to diarization changes.

Where Pith is reading between the lines

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

  • The paper does not explain the 5.7-point gap between its local dev score (23.70) and the official eval score (17.97); a plausible reading is that the local tcpMER implementation differs systematically from the official judge, so the dev-set ablation numbers should be treated as internally consistent rather than directly comparable to official results.
  • The synthetic-data count mixes dialogue groups (166,633) and utterances (67,700); the effective LoRA training set is roughly 67,700 utterance-level samples plus the dialogue groups, so the marginal contribution of each TTS pipeline remains untested.
  • A natural next experiment is to apply the same SFT+LoRA+GRPO recipe to other speech LLMs (e.g., Whisper-large-v3 or Omniasr-LLM-7B-v2) to see whether the gains are model-specific; the paper tests only Qwen3-ASR-1.7B.
  • Since SFT dominates, replacing the LoRA stage with an equal-sized SFT continuation on the synthetic data might produce a different ablation conclusion; the current design conflates adaptation method with data source.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper describes SQZ-Qwen-ASR-1.7B, a submission to MLC-SLM 2026 Challenge Task 1. The system couples a modular diarization front end (FSMN-VAD, CAMPPlus embeddings, spectral clustering with speaker count fixed to two, RTTM-based segmentation) with Qwen3-ASR-1.7B adapted in three stages: full SFT on official training data, LoRA on TTS-synthesized speech (234,333 items from three pipelines), and GRPO reinforcement learning with WER/CER-based rewards and hallucination/repetition/length penalties. On the development set the authors report an average tcpMER of 23.70 in their local scoring pipeline, compared with 30.53 for unadapted Qwen-ASR-1.7B, a 6.83-point absolute gain; on the evaluation set they report 17.97 according to the task leaderboard. Ablations show full SFT provides the largest gain (30.53 to 24.3), synthetic-speech LoRA adds 0.52, and RL adds 0.08.

Significance. If the local scoring implementation faithfully matches the official tcpMER protocol, the paper provides a clean and largely reproducible recipe whose main message (SFT on challenge data dominates; synthetic speech and RL give smaller refinements) is useful for the MLC-SLM task and for multilingual conversational ASR generally. The diarization front end and training hyperparameters are specified in unusual detail, and the per-language Table 2 allows independent checks. I do not see a circularity problem: the claim is an external benchmark measurement, not a derived quantity. The central risk is measurement validity, not internal inconsistency.

major comments (4)
  1. [Abstract; §1; §4.1; Table 2] The headline dev-set number conflates two scoring pipelines. The abstract says 'On the official development set ... 23.70', but §1 states this is 'in our local scoring pipeline' and the Table 2 caption says 'based in our local pipeline'. The 6.83-point gain (30.53 to 23.70), all ablations in Table 3, and the comparisons to Whisper/Omniasr are local; the eval number 17.97 is official/leaderboard. No evidence is given that the local tcpMER implementation (normalization, permutation search, per-language WER/CER selection, 21-condition averaging) matches the official judge. Please provide a direct local-vs-official comparison on the dev set, or report the official dev score if available, and clearly state which numbers are comparable. Without this, the abstract overstates the result as an official benchmark.
  2. [Table 2 footnote/caption] The caption says all systems are evaluated 'based in our local pipeline', but the footnote says Whisper-large-v3 and Omniasr-LLM-7B-v2 are 'evaluated under the contest protocol'. If these baseline columns are produced by the official judge while Qwen-ASR-1.7B and SQZ-Qwen-ASR-1.7B use the local pipeline, the columns are not comparable. If they are all local, the footnote is misleading. Because the local pipeline may be stricter or looser than the official protocol, cross-column absolute differences could be artifacts. Please run every system through one identical scoring implementation and state this explicitly.
  3. [§4.2; Table 3] All results are single-run numbers with no confidence intervals, significance tests, or segment/condition-level variability. The reported LoRA gain is 0.52 and the RL gain is 0.08; without some uncertainty estimate (e.g., bootstrap over the 21 language conditions, multiple decoding runs, or paired tests), these increments are not distinguishable from noise. Since §4.2 uses these small increments to support claims about 'consistent further reductions' and RL as a 'low-risk stabilizer', the absence of any variance measure is a load-bearing gap.
  4. [§4.1; §5] The dev-set score (23.70 local) differs from the eval-set score (17.97 official) by 5.7 points. If both numbers were produced under the same protocol, this large gap would suggest a substantial domain shift or model-selection effect; if they are from different pipelines, the gap is partly a scoring artifact. The paper does not discuss this discrepancy at all. Please explain the relationship between the two numbers, including whether the dev set is harder than the eval set, whether hyperparameters were selected on dev, or whether local and official scorings differ.
minor comments (4)
  1. [Abstract; §5] The abstract should carry the same qualification as §5: the dev number is 'in our local scoring pipeline'. Currently the abstract implies the 23.70 is an official dev result, which is contradicted by the body.
  2. [Table 3] Use consistent decimal precision: 24.3, 23.78, 23.70. The one-decimal 24.3 makes the LoRA increment look less precise than the others.
  3. [§2.2] The notation 'min/max/oracle num_spks=2' is unclear. State plainly whether the speaker count is always fixed to two or whether an oracle count is used for that step; the challenge setting suggests the former.
  4. [Table 2] The Urdu Qwen-ASR-1.7B value of 102.84 exceeds 100; a one-sentence explanation (e.g., tcpWER can exceed 100 after permutation/insertion penalties) would help readers interpret extreme values.

Circularity Check

0 steps flagged

No circularity found: the system is an empirical benchmark submission with an external leaderboard check, not a derivation that reduces to its own inputs.

full rationale

The paper reports an empirical SD+ASR system for the MLC-SLM 2026 challenge: a diarization front end plus Qwen3-ASR-1.7B adapted by SFT, LoRA on synthetic speech, and GRPO. The headline numbers are benchmark measurements against an external challenge protocol, not quantities derived from definitions or from fitted parameters renamed as predictions. The central 6.83-point gain is computed by comparing the adapted model with the unadapted Qwen-ASR-1.7B under the same shared diarization output (§4.1), which is a controlled experimental comparison rather than a circular reduction. The final evaluation-set tcpMER (17.97) is reported 'according to the task leaderboard' (§1, §5), providing an independent official check. The main caveat is that the development-set score (23.70) is produced by the authors' 'local scoring pipeline' (§1, §4.1), while the abstract describes it as 'on the official development set'; this is a measurement-validity and reporting concern, not circularity, because no parameter is fit to the metric and the claim remains externally testable. The ablations decompose training-stage contributions empirically; they do not assume the conclusion. No self-citation is load-bearing: references such as Qwen3-ASR, 3D-Speaker, CAMP++, and OmniVoice are external tools/models, not the authors' prior work. No uniqueness theorem, ansatz smuggled via citation, or renaming of a known result is present. Accordingly, the paper is self-contained as an empirical system description and receives a circularity score of 0.

Axiom & Free-Parameter Ledger

13 free parameters · 6 axioms · 1 invented entities

Load-bearing inputs are benchmark contracts (tcpMER definition, ~1,500 h official training data, released Qwen3-ASR base), domain assumptions about dev-eval representativeness and TTS transfer, and standard PPO/GRPO optimization math. All hyperparameters (VAD, clustering, LoRA, GRPO reward weights) are hand-tuned on the same dev set whose score is reported as the headline result, which is the main self-referential element; no new entities are postulated and nothing is derived from theory.

free parameters (13)
  • VAD posterior decision threshold = 0.5
    Hand-set; speech/non-speech boundary drives all downstream segmentation and hence every reported number.
  • VAD SNR correction terms = -0.1/0.3
    Hand-set per SNR conditions (§2.2); not described as tuned with analysis.
  • Spectral clustering cosine threshold (mer_cos) = 0.8
    Hand-set in the 3D-Speaker clustering call (§2.2).
  • Spectral clustering p-value (pval) = 0.012
    Hand-set clustering parameter (§2.2).
  • Minimum cluster size = 4
    Hand-set clustering parameter (§2.2).
  • Same-speaker merge gap = 0.8 s (disabled by default in dev)
    Post-processing threshold for merging adjacent same-speaker regions (§2.2).
  • Similarity-guided augmentation cosine threshold = 0.76–0.83
    Hand-set filtering criterion for retaining external Common Voice recordings as TTS references (§3.2, Table 1).
  • LoRA rank / alpha = r=64, alpha=128
    Chosen by hand; LoRA-ablation is not separated from the synthetic-data effect (§2.4).
  • GRPO repetition penalty weight lambda_r = 0.10
    Tuned reward weight in Eq. (1); affects the claimed RL stage behavior.
  • GRPO length-deviation penalty weight lambda_l = 0.10
    Tuned reward weight in Eq. (1).
  • Hallucination trigger thresholds = length ratio > 2.0; unique-token ratio < 0.5; 3-gram >= 2 repeats; penalty -1.0
    Hand-set rule-based penalty conditions (§2.5) that gate the largest reward swing.
  • GRPO sampling hyperparameters = K=2, temperature 0.7, top-k 20, top-p 0.95
    Hand-set; with K=2 the group-relative advantage is a binary ±0.5 sigma signal (§2.5).
  • Training schedules (SFT/LoRA/RL) = SFT 4 ep, lr 2e-5; LoRA 5 ep, lr 5e-5; RL lr 1e-6, KL beta 0.005
    Hand-set; standard values, but jointly tuned on the same dev set that is reported as the headline.
axioms (6)
  • domain assumption tcpMER as defined by the challenge organizers (WER for space-separated scripts, CER for Japanese/Korean/Thai, then unweighted average across 21 conditions) is the accepted evaluation; the paper inherits it without independent validation.
    Invoked at §1 and §4.1; all headline numbers and the 6.83-point claim are statements about this specific aggregated metric.
  • domain assumption The official development set (~4 h per language) is a representative proxy for the evaluation distribution, and model selection on it does not materially overstate generalization.
    §3.1: 'we hold out the development set for model selection'; the dev set is also the set on which the headline 23.70 is reported. The 5.7-point dev→eval drop is the observable stress on this assumption.
  • domain assumption The released Qwen3-ASR-1.7B and quoted baselines are fixed public systems, and evaluating them in the local pipeline with the same shared diarization output isolates ASR-component differences.
    §4.1 states 'The same diarization output is shared across all systems in Table 2 so that differences reflect the ASR component only.' If the released model's official scoring differs, the gain attribution changes.
  • domain assumption OmniVoice synthetic speech is acoustically and linguistically realistic enough that models trained on it transfer to real conversational audio.
    §3.2 builds a 234k-item synthesis pipeline on this premise; the ablation (+0.52 tcpMER over SFT) is the only indirect evidence.
  • domain assumption The reward R1 = 1 − WER/CER computed on normalized text is an unbiased training proxy for the final aggregated tcpMER metric.
    §2.5, Eq. (1): the GRPO objective optimizes local normalized WER/CER, while the leaderboard metric involves diarization alignment and cross-language aggregation; the paper does not analyze this mismatch.
  • standard math The clipped PPO/GRPO surrogate objective with group-relative advantages converges to a useful policy for a speech-LLM action space.
    §2.5, Eq. (2) imported from DeepSeekMath [14]; assumed to apply to ASR token-level policy optimization as in text LLMs.
invented entities (1)
  • none no independent evidence
    purpose: No new particles, forces, conserved quantities, dimensions, or theoretical constructs are introduced.
    The paper is an engineering system report; all components (VAD, embeddings, clustering, Qwen-ASR, LoRA, GRPO, OmniVoice) are cited external artifacts.

pith-pipeline@v1.3.0-alltime-deepseek · 8025 in / 19332 out tokens · 166599 ms · 2026-08-02T07:54:05.281648+00:00 · methodology

0 comments
read the original abstract

This paper describes our self-designed system for Task 1 of the MLC-SLM 2026 Challenge for multilingual two-speaker conversational speech. The system combines a modular speaker diarization front end with a challenge-adapted Qwen3-ASR-1.7B recognizer. The diarization front end performs voice activity detection, subsegment generation, CAMPPlus speaker embedding extraction, two-speaker spectral clustering, and RTTM-based audio segmentation. The resulting speaker-attributed segments are grouped by language or region and decoded by the adapted ASR model. For ASR adaptation, we first perform supervised full fine-tuning on the official training data, then apply LoRA fine-tuning with synthetic speech generated by a three-pipeline TTS-based synthetic speech augmentation framework, and finally refine the model using GRPO reinforcement learning with rewards based on WER/CER and penalties for hallucination, repetition, and length deviation. On the official development set, the full system achieves an average tcpMER of 23.70, reducing the error rate by 6.83 absolute points relative to the released Qwen-ASR-1.7B performance. On the final evaluation set, the system achieves an average tcpMER of 17.97. Ablation results show that supervised fine-tuning provides the largest gain, while synthetic-speech LoRA adaptation and reinforcement learning further improve robustness.

Figures

Figures reproduced from arXiv: 2607.08208 by Hao Wu, RongQi Han, Wei Liang, Wei Xu, Zhen Wang.

Figure 1
Figure 1. Figure 1: End-to-end framework of the SD ASR system, organized into four swim lanes. The Diarization lane performs VAD, CAMPPlus speaker embedding extraction, spectral clustering, and RTTM generation. The ASR Inference lane cuts audio according to RTTM, splits lists by language, and runs the adapted Qwen-ASR model to produce STM hypotheses. The Evaluation lane normalizes text and computes tcpMER/DER. The Model Adapt… view at source ↗
Figure 2
Figure 2. Figure 2: GRPO-based reinforcement learning for Qwen3-ASR. The speech audio and text prompt form the multimodal input; the policy model samples K = 2 candidate transcriptions, which are scored by rule-based ASR rewards (WER/CER, hallu￾cination, repetition and length penalties). The resulting group￾relative advantages Ak together with the policy probabilities π(ok) drive the GRPO actor update. weight λl = 0.10 (lengt… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 12 linked inside Pith

  1. [1]

    Introduction Transcribing two-speaker conversational recordings into speaker-attributed text is a demanding multilingual task: a single system must simultaneously decide who spoke when, segment the audio accordingly, and recognize what was said in each language. The difficulty is compounded by short turns, overlapping speech, frequent speaker switches, and...

    2026

  2. [2]

    The pipeline contains three functional parts: speaker diariza tion, au- dio segmentation and multilingual ASR, and text normalizat ion with evaluation

    System framework and training strategy Figure 1 illustrates the overall framework of our system. The pipeline contains three functional parts: speaker diariza tion, au- dio segmentation and multilingual ASR, and text normalizat ion with evaluation. The ASR model is trained separately and the n integrated into the inference pipeline. 2.1. Stage-wise infere...

    Pith/arXiv arXiv 2026

  3. [3]

    Dataset 3.1. Training data The MLC-SLM 2026 Task 1 training set provides two-speaker conversational speech recorded at 16 kHz in quiet indoor en- vironments, with oracle segmentation and speaker labels [ 1]. It covers 21 language/region conditions, built on the origi nal 11-language release (English ≈500 h, other languages ≈100 h each, totaling ≈1,500 h) ...

    2026

  4. [4]

    Results We report tcpMER on the official development set under the challenge evaluation protocol

    Experiments 4.1. Results We report tcpMER on the official development set under the challenge evaluation protocol. For ASR decoding we use greedy decoding at temperature 0 with a maximum output length of 256 tokens; diarization uses CAMPPlus embeddings and spectral clustering with the speaker count fixed to two, f ol- lowed by RTTM-driven segmentation. The ...

  5. [5]

    Conclusion We present an end-to-end speaker-diarization and ASR syste m for the MLC-SLM 2026 Task 1 challenge, built on Qwen3- ASR-1.7B and adapted through full supervised fine-tuning, synthetic-speech LoRA, and GRPO-based reinforcement lear n- ing. A three-pipeline TTS-based synthetic speech augmenta - tion framework based on OmniV oice contributes 234,33...

    2026

  6. [6]

    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,” in Proceedings of INTERSPEECH, 2023

    2023

  7. [7]

    The Multilingual Conversational Speech Language Model Challenge 2026,

    MLC-SLM 2026 Challenge Organizers, “The Multilingual Conversational Speech Language Model Challenge 2026,” https://www.nexdata.ai/competition/mlc-slm, 2026, accessed: 2026-07-06

    2026

  8. [8]

    3D-Speaker- Toolkit: An open-source toolkit for multimodal speaker ver ifica- tion and diarization,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, T. Zhu, R. Huang, C. Deng, Q. Chen, S. Zhang, W. Wang, and X. Li, “3D-Speaker- Toolkit: An open-source toolkit for multimodal speaker ver ifica- tion and diarization,” arXiv preprint arXiv:2403.19971, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Qwen3-ASR technical report,

    X. Shi, X. Wang, Z. Guo, Y . Wang, P . Zhang, X. Zhang, Z. Guo, H. Hao, Y . Xi, B. Y ang, J. Xu, J. Zhou, and J. Lin, “Qwen3-ASR technical report,” arXiv preprint arXiv:2601.21337, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    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,” in International Conference on Learning Repre- sentations (ICLR), 2022

    2022

  11. [11]

    OmniV oice: Towards om- nilingual zero-shot text-to-speech with diffusion langua ge mod- els,

    H. Zhu, L. Y e, W. Kang, Z. Y ao, L. Guo, F. Kuang, Z. Han, W. Zhuang, L. Lin, and D. Povey, “OmniV oice: Towards om- nilingual zero-shot text-to-speech with diffusion langua ge mod- els,” arXiv preprint arXiv:2604.00688, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Qwen2 technical report,

    A. Y ang, B. Y ang, B. Hui, B. Zheng, B. Y u, C. Zhou, C. Li, C. Li, D. Liu, F. Huang et al., “Qwen2 technical report,” arXiv preprint arXiv:2407.10671, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    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,” in IEEE Interna- tional Conference on Acoustics, Speech and Signal Processi ng (ICASSP), 2023

    2023

  14. [14]

    FunASR: A fundamental end-to- end speech recognition toolkit,

    Z. Gao, Z. Li, J. Wang, H. Luo, X. Shi, M. Chen, Y . Li, L. Zuo, Z. Du, Z. Xiao, and S. Zhang, “FunASR: A fundamental end-to- end speech recognition toolkit,” arXiv preprint arXiv:2305.11013, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    End-to-end neural speaker diarization wi th permutation-free objectives,

    Y . Fujita, N. Kanda, S. Horiguchi, Y . Xue, K. Nagamatsu, and S. Watanabe, “End-to-end neural speaker diarization wi th permutation-free objectives,” in Proceedings of INTERSPEECH , 2019

    2019

  16. [16]

    Meeting recognition w ith continuous speech separation and transcription-supported diariza- tion,

    T. von Neumann and K. Kinoshita, “Meeting recognition w ith continuous speech separation and transcription-supported diariza- tion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024, arXiv:2309.16482

    Pith/arXiv arXiv 2024

  17. [17]

    Qwen2-Audio technical re - port,

    Y . Chu, J. Xu, Q. Y ang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv , J. He, J. Lin, C. Zhou, and J. Zhou, “Qwen2-Audio technical re - port,” arXiv preprint arXiv:2407.10759, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    MLC-SLM 2026 Task 1 Official Baseline,

    MLC-SLM 2026 Challenge Organizers, “MLC-SLM 2026 Task 1 Official Baseline,” https://github.com/alanshaoTT/MLC-SLM-2nd-Task1-Baseline, 2026, accessed: 2026-07-06

    2026

  19. [19]

    Qwen3 technical report,

    A. Y ang, A. Li, B. Y ang, B. Zhang, B. Hui, B. Zheng, B. Y uet al., “Qwen3 technical report,” arXiv preprint arXiv:2505.09388 , 2025

    Pith/arXiv arXiv 2025

  20. [20]

    DeepSeekMath: Pushing the limits of mathematical reasoning in open language model s,

    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 model s,” arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Text-to-speech data augmentation for lo w resource speech recognition,

    R. Zevallos, “Text-to-speech data augmentation for lo w resource speech recognition,” arXiv preprint arXiv:2204.00291, 2022

    Pith/arXiv arXiv 2022

  22. [22]

    Frustratingly easy data augm entation for low-resource ASR,

    K. Ibaraki and D. Chiang, “Frustratingly easy data augm entation for low-resource ASR,” arXiv preprint arXiv:2509.15373, 2025

    arXiv 2025

  23. [23]

    Com - mon V oice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. H en- retty, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Com - mon V oice: A massively-multilingual speech corpus,” inProceed- ings of the Twelfth Language Resources and Evaluation Confe r- ence, 2020, pp. 4218–4222

    2020

  24. [25]

    ASR er- ror correction using large language models,

    R. Ma, M. Qian, M. J. F. Gales, and K. Knill, “ASR er- ror correction using large language models,” arXiv preprint arXiv:2409.09554, 2024

    Pith/arXiv arXiv 2024

  25. [512]

    This stage serves as the final ASR refinement after supervised fine-tuning and LoRA adaptation

    A low-variance KL loss regularizes the policy toward th e reference with β = 0 .005; KL is not injected into the reward. This stage serves as the final ASR refinement after supervised fine-tuning and LoRA adaptation