Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Weakly Supervised Data Refinement and Flexible Sequence Compression for Efficient Thai LLM-based ASR

T0 review · 5 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 This paper claims the first LLM-based Thai ASR system, EThai-ASR, and reports state-of-the-art character error rates on several Thai test sets through self-evolving weak-label refinement and pluggable sequence compression.

desk verdict This is a useful Thai LLM-ASR system with a valuable dataset release, but the SOTA and <5% degradation claims outrun the evidence as written; the trainable compression modes are the real contribution. read the letter →

arxiv 2505.22063 v1 pith:A5YDA6LL submitted 2025-05-28 cs.SD eess.AS

Mingchen Shao , Xinfa Zhu , Chengyou Wang , Bingshen Mu , Hai Li , Ying Yan , Junhui Liu , Danming Xie
show 1 more author
Lei Xie
This is my paper
classification cs.SDeess.AS
keywords ThaiASRLLM-basedspeechrecognitionweaklabelrefinementself-evolvingdatasequencecompressionZipformerlow-resourcecosinesimilarity
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

EThai-ASR aims to make Thai speech recognition both accurate and practical despite scarce precisely annotated data and heavy computational demands. The paper claims this is the first system to use a Thai large language model as the ASR decoder, pairing a Zipformer speech encoder with a Thai LLM. To overcome data scarcity, it introduces a self-evolving refinement loop that cleans weak web-crawled labels by keeping only labels on which the encoder and a Whisper-based anchor model agree. To reduce compute, it adds a pluggable sequence compression module that drops redundant adjacent speech frames based on cosine similarity, achieving 1.5x to 2.1x acceleration with less than 5% accuracy loss in trainable modes. If correct, the system establishes a practical template for low-resource LLM-based ASR and releases 16k hours of refined Thai transcripts to the community.

What carries the argument

The load-bearing mechanism is the self-evolving data refinement loop: a pretrained Zipformer encoder is trained on a hybrid set of 500 hours of precise labels (CommonVoice Thai) and 16k hours of weak labels (Gigaspeech2 and MSR-86k), then predicts labels, and only samples whose edit distance to labels from a Monsoon-whisper anchor model falls below a threshold are kept as refined labels, with the process iterated three times. This yields Zipformer-iter3, which then feeds a 4-layer Transformer adapter and linear projector into a Typhoon2 Thai LLM, with LoRA finetuning in the second training stage. The efficiency claim rests on frame-level cosine similarity, $\text{CosineSim}(x_t, x_{t+1}) = \frac{x_t \cdot x_{t+1}}{\|x_t\|\|x_{t+1}\|}$, where adjacent frames above a threshold $\theta$ are deemed redundant and the later frame is removed before the LLM sees the sequence.

What would settle it

Take a human-transcribed Thai sample, run Algorithm 1 with the Monsoon-whisper anchor, and compare the edit-distance-filtered labels against the human transcripts; if the surviving labels have no lower error than the original weak labels, the self-evolving refinement is selecting for model consensus rather than correctness.

Watch

Extended reading notes

Core claim

EThai-ASR is presented as the first LLM-based Thai ASR system, integrating a Zipformer speech encoder refined through three iterations of weak-label correction with a Typhoon2 Thai LLM decoder via a Transformer adapter and linear projector. The paper reports state-of-the-art CER on the Gigaspeech2 Thai test set (11.67%), CommonVoice Thai test (3.11%), and FLEUR Thai test (9.36%), surpassing Whisper-large-v2/v3, Monsoon-whisper-medium, and Zipformer baselines. A second central claim is that its cosine-similarity-based sequence compression module, which removes redundant adjacent encoder frames, can accelerate training and inference by 1.5x to 2.1x while keeping accuracy degradation below 5% when integrated through from-scratch training or efficient finetuning of the linear projector alone.

Load-bearing premise

The refinement loop assumes that when the Zipformer model and the Monsoon-whisper anchor agree on a Thai label, that label is correct; if the anchor is systematically wrong for Thai tones, proper nouns, or dialects, the retained labels inherit and reinforce those errors.

Editorial extensions

If this is right

  • EThai-ASR reports SOTA CER of 11.67% on Gigaspeech2 Thai, 3.11% on CommonVoice Thai, and 9.36% on the out-of-domain FLEUR Thai test set.
  • The self-evolving refinement improves average CER from 9.05% (baseline) to 8.58% over three iterations, with relative gains on individual sets ranging from about 3% to 17%.
  • The pluggable sequence compression module yields 1.5x to 2.1x speedup with under 5% CER degradation in from-scratch and efficient-finetuning modes across three LLM backbones.
  • Efficient finetuning updates only the linear projector, roughly 3M parameters, using a few hours of data and about 30 minutes on eight Nvidia 3090 GPUs.
  • The refined text transcripts for Gigaspeech2 and MSR-86k, totaling 16k hours of Thai speech data, are released publicly.

Reading between the lines

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

  • The open-sourced refined transcripts could serve as a pretraining or distillation corpus for other Thai speech models, not only LLM-based ASR systems.
  • Because the compression threshold is a tunable scalar, the same module could be adapted per domain or per latency budget, though the paper only reports a fixed balanced point near $\theta=0.887$.
  • The compression mechanism operates purely on encoder output, so it may transfer to other encoder-decoder LLM-based ASR systems beyond Thai, provided the linear projector is lightly finetuned.
  • If the Monsoon-whisper anchor is systematically biased on Thai tones, proper nouns, or dialects, the agreement filter could consolidate those biases into the refined labels; the paper does not evaluate label quality against human transcripts.
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

5 major / 5 minor

Summary. The paper presents EThai-ASR, a Thai ASR system that combines a Zipformer speech encoder with a Typhoon2 Thai LLM decoder through a Transformer adapter and linear projector. Its three contributions are: a self-evolving data refinement strategy that iteratively corrects weak labels using agreement between a Zipformer model and a Monsoon-whisper anchor via edit-distance filtering; a two-stage training procedure for aligning speech features with the LLM semantic space; and a pluggable sequence compression module that removes similar adjacent frames based on cosine similarity, with three integration modes. Experiments report CER on Gigaspeech2, CommonVoice, and FLEUR Thai test sets, comparing against Zipformer and Whisper baselines, and the authors release refined transcripts.

Significance. If the claims are substantiated, the paper would provide a useful low-resource Thai ASR system, an open set of refined labels for 16k hours of Thai speech, and a pluggable compression technique that can accelerate LLM-based ASR without retraining in some modes. The evaluation covers in-domain and out-of-domain test sets and three different LLM backbones, and the release of refined labels is a concrete community contribution. However, the headline claims of state-of-the-art accuracy and of less than 5% accuracy degradation under compression are not fully supported by the reported experiments, so the significance of the contribution cannot be assessed at face value.

major comments (5)
  1. [Abstract and Section 1; Table 3] The claim that the sequence compression module achieves '1.5x to 2.1x acceleration while incurring an accuracy degradation of less than 5%' is contradicted by the inference-only rows of Table 3. For Typhoon2-Llama3.1-8B, inference-only at theta=0.95 raises CER from 11.67 to 14.10, a 20.8% relative increase, and for Typhoon-7B it raises CER from 16.87 to 27.31, a 61.9% relative increase. Even interpreting 'less than 5%' as absolute CER points, the 7B result is not covered. The claim should be restricted to the from-scratch and efficient-finetuning modes, or the paper should report relative degradation consistently for all modes.
  2. [Sections 1 and 3.3; Table 2] The 'state-of-the-art accuracy' claim is not established because Table 2 omits Thai-specific baselines from the paper's own reference [14], 'Comprehensive Benchmarking and Analysis of Open Pretrained Thai Speech Recognition Models.' The table only compares Zipformer-gigaspeech2, Whisper-large-v2/v3, and Monsoon-whisper-medium; no CERs for any models benchmarked in [14] are reported, nor are any prior Thai LLM-based ASR systems compared. Without these comparisons, the SOTA claim is unsupported. Similarly, the 'first to apply LLMs to Thai ASR' claim needs a literature search or at least a tabulated prior-work survey.
  3. [Algorithm 1 and Section 2.1] The self-evolving data refinement presumes that agreement between the Zipformer model and the Monsoon-whisper-medium anchor implies a correct label. If the anchor has systematic Thai-specific errors, for example on tones, proper nouns, or dialects, then iterative training on the model's own predictions can consolidate those errors rather than remove them. The paper provides no analysis of the anchor's error profile or any human evaluation of retained versus discarded labels, so this assumption is load-bearing for the data-refinement contribution and should be tested on a sampled subset.
  4. [Tables 1 and 2] The reported CER improvements are very small and are given without error bars or significance tests. For example, Table 1 shows absolute improvements of 0.38-0.69 CER across refinement iterations, and Table 2 shows EThai-ASR versus Zipformer-iter3 differences of 0.41, 0.35, and 0.85 on the three test sets. Because CER on finite test sets has sampling variability, these differences may be within noise. The paper should at least report test-set sizes and confidence intervals, and ideally multiple runs, before claiming that the refinement and the LLM integration reliably improve accuracy.
  5. [Section 3.4.1 and Figure 2] The 'CER Retention' metric used to justify the threshold of 0.887 is never defined in the text, and the dataset on which the threshold-performance trade-off is measured is not stated. If the thresholds were selected on the same evaluation sets used in Table 3, then the reported trade-off curve has a tuning-on-test component. The paper should define CER Retention, specify the development/test split, and clarify how thresholds were chosen.
minor comments (5)
  1. [References] Reference [2], cited for LLama in the introduction, is actually a speech emotion recognition paper and appears to be an incorrect citation for the Llama model.
  2. [Section 3.3 heading] The heading 'Evaluation of LLM-baed ASR' contains a typo: it should be 'LLM-based ASR'.
  3. [Table 3] The model name 'Typhoon-7B' is inconsistent with the 'Typhoon2-...' names used for the other two backbones; please align the terminology.
  4. [Table 3 and Section 3.4] Table 3 reports CER for each compression mode but does not report the corresponding Speedup Ratio values, so the paper's 1.5x-2.1x acceleration claim cannot be traced to the table. Please add SR per mode or point to where they are measured.
  5. [Section 2.3.2 and Section 3.1] Several experimental details needed for reproducibility are missing, including batch size, optimizer, training epochs, warmup, LoRA target modules, adapter dimensions, and the exact amount of data used for efficient finetuning.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the reported CER results are held-out test measurements and the method is fully specified in Algorithm 1 and Section 2.3.

full rationale

The paper's central claims are evaluated on held-out test sets (Giga2 Test, CV Test, FLEUR Test) against external baselines including Whisper variants, Monsoon-whisper, and Zipformer-gigaspeech2, so the reported accuracy numbers are measurements rather than quantities forced by construction. The data-refinement algorithm is explicit: a Zipformer model is fine-tuned on a hybrid set, its predictions are kept when they agree with a Monsoon-whisper anchor within an edit-distance threshold of 10, and the filtered predictions are used as training labels for the next iteration; the resulting Zipformer-iter3 is then evaluated independently. No equation defines the outcome in terms of the inputs, and the edit-distance threshold is a stated hyperparameter rather than a fit to the headline test numbers. The cosine-similarity compression module is likewise an explicit preprocessing rule; the thresholds 0.87 and 0.95 are chosen to match target compression ratios, and the CER values in Table 3 are measured results, not predictions derived from the threshold formula. The self-citations in the paper ([22], [28]) are not load-bearing: Algorithm 1 and the two-stage training description stand on their own, and no cited uniqueness theorem or prior result is used to forbid alternatives or to define the architecture. The SOTA claim's reliance on a narrow baseline set is an external-validity or completeness concern, not a circularity, and therefore does not raise the circularity score.

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

The central empirical claims rest on two hand-set thresholds (edit distance 10, cosine similarity around 0.87 to 0.95) plus standard domain assumptions about label reliability and frame redundancy. No new physical or model entities are introduced.

free parameters (5)
  • edit_distance_threshold_tau = 10
    Samples with edit distance greater than 10 between Zipformer and whisper anchor predictions are discarded (Section 2.1, Algorithm 1). The value is set by the authors without justification from data outside the test set.
  • cosine_similarity_threshold_theta = 0.87 / 0.887 / 0.95
    Used to prune redundant frames (Section 3.4). The value 0.887 is called the balance point from Figure 2; 0.87 and 0.95 are chosen to hit 25 percent and 50 percent compression targets on evaluation data.
  • refinement_iterations_n = 3
    Algorithm 1 runs three iterations; performance at iteration 3 is reported as Zipformer-iter3 (Table 1).
  • lora_rank_alpha = rank 32, alpha 16
    Set in Section 3.3 for second-stage LoRA fine-tuning.
  • learning_rate = 5e-4
    Used in both training stages (Section 3.3).
assumptions (4)
  • domain assumption Whisper-anchor labels are a reliable correctness proxy for Thai transcription.
    Algorithm 1 uses edit distance to Monsoon-whisper-medium-gigaspeech2 predictions to decide which Zipformer predictions become refined labels. If the anchor is biased on Thai tones or proper nouns, the filtered labels inherit that bias.
  • domain assumption Adjacent encoder frames with cosine similarity above theta are redundant for downstream LM decoding.
    Section 2.3.1 assumes frame-wise cosine similarity identifies removable frames without losing phonetically critical details in Thai, which has short-vowel and tonal contrasts.
  • domain assumption Test sets are disjoint from the training and refinement data.
    The paper never explicitly states that the Giga2 Test and CV Test partitions are excluded from the 13k plus 0.5k hours used for refinement, though this is standard practice.
  • standard math Cosine similarity and edit distance are standard metrics used without modification.
    Used in Equation 1 and Algorithm 1 without derivation, treated as standard definitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weakly Supervised Data Refinement and Flexible Sequence Compression for Efficient Thai LLM-based ASR." pith.science (2026). https://pith.science/paper/A5YDA6LL

@misc{pith2026250522063,
  author       = {Pith},
  title        = {Pith review of: Weakly Supervised Data Refinement and Flexible Sequence Compression for Efficient Thai LLM-based ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A5YDA6LL}},
  note         = {Machine review of arXiv:2505.22063}
}
read the original abstract

Despite remarkable achievements, automatic speech recognition (ASR) in low-resource scenarios still faces two challenges: high-quality data scarcity and high computational demands. This paper proposes EThai-ASR, the first to apply large language models (LLMs) to Thai ASR and create an efficient LLM-based ASR system. EThai-ASR comprises a speech encoder, a connection module and a Thai LLM decoder. To address the data scarcity and obtain a powerful speech encoder, EThai-ASR introduces a self-evolving data refinement strategy to refine weak labels, yielding an enhanced speech encoder. Moreover, we propose a pluggable sequence compression module used in the connection module with three modes designed to reduce the sequence length, thus decreasing computational demands while maintaining decent performance. Extensive experiments demonstrate that EThai-ASR has achieved state-of-the-art accuracy in multiple datasets. We release our refined text transcripts to promote further research.

Figures

Figures reproduced from arXiv: 2505.22063 by the authors.

Figure 1
Figure 1. The architecture of EThai-ASR. The left panel illustrates redundant speech frames are dynamically removed via cosine similarity to reduce sequence length before LLM processing. The right panel illustrates three integration modes: in from-scratch training, the Transformer adapter, linear projector, and LLM LoRA are finetuned after redundancy removal; efficient finetuning adjusts only the linear projector for comparab… view at source ↗
Figure 2
Figure 2. Impact of cosine similarity thresholds on SR and CER retention. The red line represents the variation in SR, while the blue line illustrates the changes in CER Retention. 3.4.1. Threshold-performance-acceleration tradeoff We evaluate the impact of the cosine similarity threshold θ on both the SR and CER retention [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. WenetSpeech-Yue: A Large-scale Cantonese Speech Corpus with Multi-dimensional Annotation

    cs.SD 2025-09 conditional novelty 6.0 of 10

    The authors built and released the largest open-source Cantonese speech corpus (21,800 hours, 10 domains, rich metadata), and show that models trained on it match or beat existing speech recognition and synthesis systems.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [14]

    GigaSpeech 2: An Evolving, Large- Scale and Multi-domain ASR Corpus for Low-Resource Lan- guages with Automated Crawling, Transcription and Refinement,

    Y . Yang, Z. Song, J. Zhuo, M. Cui, J. Li, B. Yang, Y . Du, Z. Ma, X. Liu, Z. Wanget al., “GigaSpeech 2: An Evolving, Large- Scale and Multi-domain ASR Corpus for Low-Resource Lan- guages with Automated Crawling, Transcription and Refinement,” arXiv preprint arXiv:2406.11546, 2024

    arXiv 2024

  2. [1]

    However, in low-resource settings, limited annotated data and computational resources hinder performance, particularly in Thai ASR

    Introduction Recent advances in automatic speech recognition (ASR) have led to significant improvements in high-resource scenarios. However, in low-resource settings, limited annotated data and computational resources hinder performance, particularly in Thai ASR. Meanwhile, large language models (LLMs), such as GPT [1] and LLama [2], have demonstrated imp...

    arXiv 2025

  3. [2]

    To achieve this goal, EThai-ASR first introduces a self-evolving data refine- ment strategy to improve the quality of weakly labeled datasets

    Method EThai-ASR aims to leverage prior knowledge from text LLM for low-resource Thai ASR and address two challenges: data scarcity and high computational demands. To achieve this goal, EThai-ASR first introduces a self-evolving data refine- ment strategy to improve the quality of weakly labeled datasets. Subsequently, EThai-ASR is trained on the refined ...

    work page

  4. [3]

    Giga2 Test

    Experiment 3.1. Experiment setup Our experiments employ a combination of weak labels and pre- cise labels. Specifically, the weak labels data consist of Gi- gaspeech2 (13k hours) and MSR-86k (3k hours), while the pre- cise labels dataset is CommonV oice Thai (500 hours). We eval- uate our models on three test sets: two in-domain sets (Gi- gaspeech2 Test a...

    work page

  5. [4]

    Conclusion In this work, we propose EThai-ASR for low-resource scenar- ios. We introduce a self-evolving data refinement strategy that improves weak labels, enhances recognition accuracy by pro- jecting speech features into a Thai LLM’s semantic space using a two-stage training strategy, and reduces computational over- head through threshold-aware sequenc...

    work page

  6. [5]

    Gpt-4 technical report,

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

    arXiv 2023

  7. [6]

    Enhanc- ing Speech Emotion Recognition Using Dual Feature Extraction Encoders,

    I. Pulatov, R. Oteniyazov, F. Makhmudov, and Y . Cho, “Enhanc- ing Speech Emotion Recognition Using Dual Feature Extraction Encoders,”Sensors, vol. 23, no. 14, p. 6640, 2023

    work page 2023

  8. [7]

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

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

    arXiv 2023

Show all 40 references
  1. [8]

    Qwen2-audio techni- cal report,

    Chu, Yunfei and Xu, Jin and Yang, Qian and Wei, Haojie and Wei, Xipin and Guo, Zhifang and Leng, Yichong and Lv, Yuanjun and He, Jinzheng and Lin, Junyang and others, “Qwen2-audio techni- cal report,”arXiv preprint arXiv:2407.10759, 2024

    2024 arXiv

  2. [9]

    SALMONN: Towards Generic Hearing Abilities for Large Language Models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “SALMONN: Towards Generic Hearing Abilities for Large Language Models,” inProc. ICLR, 2024

    2024

  3. [10]

    Meta- Adaptable-Adapter: Efficient adaptation of self-supervised mod- els for low-resource speech recognition,

    Y . Chen, H. Zhang, X. Yang, W. Zhang, and D. Qu, “Meta- Adaptable-Adapter: Efficient adaptation of self-supervised mod- els for low-resource speech recognition,”Neurocomputing, vol. 609, p. 128493, 2024

    2024

  4. [11]

    Stable Distilla- tion: Regularizing Continued Pre-Training for Low-Resource Au- tomatic Speech Recognition,

    A. Seth, S. Ghosh, S. Umesh, and D. Manocha, “Stable Distilla- tion: Regularizing Continued Pre-Training for Low-Resource Au- tomatic Speech Recognition,” inProc. ICASSP, 2024, pp. 10 821– 10 825

    2024

  5. [12]

    Com- mon V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Hen- retty, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Com- mon V oice: A Massively-Multilingual Speech Corpus,” inProc. LREC, 2020, pp. 4218–4222

    2020

  6. [13]

    Lib- rispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An ASR corpus based on public domain audio books,” inProc. ICASSP 2015, pp. 5206–5210

    2015

  7. [15]

    MSR- 86K: An Evolving, Multilingual Corpus with 86,300 Hours of Transcribed Audio for Speech Recognition Research,

    S. Li, Y . You, X. Wang, Z. Tian, K. Ding, and G. Wan, “MSR- 86K: An Evolving, Multilingual Corpus with 86,300 Hours of Transcribed Audio for Speech Recognition Research,”arXiv preprint arXiv:2406.18301, 2024

    2024 arXiv

  8. [16]

    Thai-Dialect: Low Resource Thai Dialectal Speech to Text Cor- pora,

    A. Suwanbandit, J. Chitiyaphol, S. Chuenchom, K. Kwiecien, H. Sawal, R. Uthai, O. Sangpetch, and E. Chuangsuwanich, “Thai-Dialect: Low Resource Thai Dialectal Speech to Text Cor- pora,” inProc. ASRU 2023, pp. 1–8

    2023

  9. [17]

    Automatic Speech Recognition Techniques for Transcription of Thai Traditional Medicine Texts,

    J. Popun, W. Lee, and A. Prayote, “Automatic Speech Recognition Techniques for Transcription of Thai Traditional Medicine Texts,” inProc. ISOCC 2024, pp. 256–257

    2024

  10. [18]

    Comprehensive Benchmarking and Analysis of Open Pretrained Thai Speech Recognition Models,

    P. Tipakasorn, O. Chatthong, R. Yonehana, and K. Thangthai, “Comprehensive Benchmarking and Analysis of Open Pretrained Thai Speech Recognition Models,” inProc. O-COCOSDA 2024, pp. 1–7

    2024

  11. [19]

    Typhoon: Thai large language models,

    K. Pipatanakul, P. Jirabovonvisut, P. Manakul, S. Sripaisarn- mongkol, R. Patomwong, P. Chokchainant, and K. Tharnpip- itchai, “Typhoon: Thai large language models,” arXiv preprint arXiv:2312.13951,2023

    2023 arXiv

  12. [20]

    Typhoon 2: A Fam- ily of Open Text and Multimodal Thai Large Language Models,

    K. Pipatanakul, P. Manakul, N. Nitarach, W. Sirichotedumrong, S. Nonesung, T. Jaknamon, P. Pengpun, P. Taveekitworachai, A. Na-Thalang, S. Sripaisarnmongkolet al., “Typhoon 2: A Fam- ily of Open Text and Multimodal Thai Large Language Models,” arXiv preprint arXiv:2412.13702, 2024

    2024 arXiv

  13. [21]

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

    Y . Bai, J. Chen, J. Chen, W. Chen, Z. Chen, C. Ding, L. Dong, Q. Dong, Y . Du, and K. G. et al., “Seed-asr: Understanding di- verse speech and contexts with llm-based speech recognition,” arXiv preprint arXiv:2407.04675, 2024

    2024 arXiv

  14. [22]

    Audiopalm: A large language model that can speak and listen,

    P. K. Rubenstein, C. Asawaroengchai, D. D. Nguyen, A. Bapna, Z. Borsos, F. d. C. Quitry, P. Chen, D. E. Badawy, W. Han, E. Kharitonovet al., “Audiopalm: A large language model that can speak and listen,”arXiv preprint arXiv:2306.12925, 2023

    2023 arXiv

  15. [23]

    Enhancing low-resource language and in- struction following capabilities of audio language models,

    P. Manakul, G. Sun, W. Sirichotedumrong, K. Tharnpipitchai, and K. Pipatanakul, “Enhancing low-resource language and in- struction following capabilities of audio language models,”arXiv preprint arXiv:2409.10999, 2024

    2024 arXiv

  16. [24]

    Skipformer: A Skip-and-Recover Strategy for Efficient Speech Recognition,

    W. Zhu, S. Sun, C. Shan, P. Fan, and Q. Yang, “Skipformer: A Skip-and-Recover Strategy for Efficient Speech Recognition,” in Proc. ICME 2024

    2024

  17. [25]

    Zipformer: A faster and better encoder for automatic speech recognition,

    Z. Yao, L. Guo, X. Yang, W. Kang, F. Kuang, Y . Yang, Z. Jin, L. Lin, and D. Povey, “Zipformer: A faster and better encoder for automatic speech recognition,” inProc. ICLR 2024

    2024

  18. [26]

    Cosyaudio: Improving audio generation with confidence scores and synthetic captions,

    X. Zhu, W. Tian, X. Wang, L. He, X. Wang, S. Zhao, and L. Xie, “Cosyaudio: Improving audio generation with confidence scores and synthetic captions,”arXiv preprint arXiv:2501.16761, 2025

    2025 arXiv

  19. [27]

    scPLAN: a hierar- chical computational framework for single transcriptomics data annotation, integration and cell-type label refinement,

    Q. Guo, M. Yuan, L. Zhang, and M. Deng, “scPLAN: a hierar- chical computational framework for single transcriptomics data annotation, integration and cell-type label refinement,”Briefings in Bioinformatics, vol. 25, no. 4, p. bbae305, 2024

    2024

  20. [28]

    Data driven multi-objective economic-environmental robust optimiza- tion for refinery planning with multiple modes under uncertainty,

    J. Long, N. Wang, J. Zhai, C. Liang, S. Jiang, and L. Zhao, “Data driven multi-objective economic-environmental robust optimiza- tion for refinery planning with multiple modes under uncertainty,” Computers & Industrial Engineering, vol. 198, p. 110697, 2024

    2024

  21. [29]

    A survey of dataset refinement for problems in computer vision datasets,

    Z. Wan, Z. Wang, C. Chung, and Z. Wang, “A survey of dataset refinement for problems in computer vision datasets,”ACM com- puting surveys, vol. 56, no. 7, pp. 1–34, 2024

    2024

  22. [30]

    A multimodal cross- transformer-based model to predict mild cognitive impairment us- ing speech, language and vision,

    F. F. Poor, H. H. Dodge, and M. H. Mahoor, “A multimodal cross- transformer-based model to predict mild cognitive impairment us- ing speech, language and vision,”Comput. Biol. Medicine, vol. 182, p. 109199, 2024

    2024

  23. [31]

    Qformer: An efficient quaternion transformer for image denoising,

    B. Jiang, Y . Lu, G. Lu, and B. Zhang, “Qformer: An efficient quaternion transformer for image denoising,” inProc. IJCAI 2024

    2024

  24. [32]

    Unveiling the Potential of LLM-Based ASR on Chinese Open-Source Datasets,

    X. Geng, T. Xu, K. Wei, B. Mu, H. Xue, H. Wang, Y . Li, P. Guo, Y . Dai, L. Li, M. Shao, and L. Xie, “Unveiling the Potential of LLM-Based ASR on Chinese Open-Source Datasets,” inProc. ISCSLP 2024, pp. 26–30

    2024

  25. [33]

    LoRA: Low-Rank Adaptation of Large Language 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 Language Models,” inProc. ICLR 2022

    2022

  26. [34]

    Lora+: Efficient low rank adap- tation of large models,

    S. Hayou, N. Ghosh, and B. Yu, “Lora+: Efficient low rank adap- tation of large models,” inProc. ICML 2024

    2024

  27. [35]

    Sequence-level Knowledge Distillation for Model Compression of Attention- based Sequence-to-sequence Speech Recognition,

    R. M. Mun’im, N. Inoue, and K. Shinoda, “Sequence-level Knowledge Distillation for Model Compression of Attention- based Sequence-to-sequence Speech Recognition,” inProc. ICASSP 2019

    2019

  28. [36]

    Once-for-all sequence compres- sion for self-supervised speech models,

    H. Chen, Y . Meng, and H. Lee, “Once-for-all sequence compres- sion for self-supervised speech models,” inProc. ICASSP 2023, pp. 1–5

    2023

  29. [37]

    Visionzip: Longer is better but not necessary in vision language models,

    S. Yang, Y . Chen, Z. Tian, C. Wang, J. Li, B. Yu, and J. Jia, “Visionzip: Longer is better but not necessary in vision language models,”arXiv preprint arXiv:2412.04467, 2024

    2024

  30. [38]

    VideoEspresso: A Large-Scale Chain- of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection,

    S. Han, W. Huang, H. Shi, L. Zhuo, X. Su, S. Zhang, X. Zhou, X. Qi, Y . Liao, and S. Liu, “VideoEspresso: A Large-Scale Chain- of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection,”arXiv preprint arXiv:2411.14794, 2024

    2024 arXiv

  31. [39]

    FLEURS: FEW-Shot Learning Evaluation of Universal Representations of Speech,

    A. Conneau, M. Ma, S. Khanuja, Y . Zhang, V . Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna, “FLEURS: FEW-Shot Learning Evaluation of Universal Representations of Speech,” inProc. SLT 2022, pp. 798–805

    2022

  32. [40]

    Robust Speech Recognition via Large-Scale Weak Supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,” inProc, ICML 2023, pp. 28 492–28 518

    2023

Pith tools

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