Pith. sign in

REVIEW 3 major objections 5 minor 42 references

The TEA-ASLP System for Multilingual Conversational Speech Recognition and Speech Diarization in MLC-SLM 2025 Challenge

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that routing per-language LoRA adapters by known language ID and adding CTC-predicted tokens as prompts yields a 30.8% relative WER reduction on multilingual conversational ASR, with 9.60% WER on the official test set.

desk verdict Solid challenge system with real leaderboard results, but the abstract pairs the wrong number with the final system and the mLoRA contribution is not isolated. read the letter →

arxiv 2507.18051 v1 pith:26SV2GCA submitted 2025-07-24 cs.SD eess.AS

Hongfei Xue , Kaixun Huang , Zhikai Zhou , Shen Huang , Shidong Shang This is my paper · ORCID
classification cs.SDeess.AS
keywords multilingualautomaticspeechrecognitionlargelanguagemodelidentificationmixture-of-expertsLoRACTCpromptingspeakerdiarizationdualencoderfusionMLC-SLM2025
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 is a system report from the MLC-SLM 2025 Challenge, aiming to show that a speech large language model can be made substantially better at multilingual conversational ASR when the known language of each utterance is used as a routing signal. The authors claim that three additions to their prior Ideal-LLM architecture — dual Whisper/MMS encoders fused by a language-adapted connector, a language-ID-routed mixture-of-experts LoRA adapter, and CTC-predicted tokens fed into the LLM as prompts — reduce average word error rate by 30.8% relative to the challenge baseline. With roughly 180k hours of training data across 11 languages, their system reaches a final WER of 9.60% on Task I and a time-constrained minimum-permutation WER of 17.49% on Task II, which placed first and second. A sympathetic reader would care because the paper gives a concrete, reproducible recipe for conditioning multilingual speech LLMs on language identity, and because the same machinery is shown to help diarized ASR when the speaker-diarization front end is specialized to a single language.

What carries the argument

The load-bearing mechanism is the LID-routed multilingual MoE LoRA (mLoRA) combined with the language-adapted connector. mLoRA assigns each of the 11 languages its own low-rank adapter on both the dual encoders and the Qwen3-8B decoder, and the known language ID selects which adapter is active, letting the shared base model specialize its weights per language without a full fine-tune. The connector performs a weighted fusion of Whisper and MMS representations with per-language sigmoid weights, also selected by LID, before projecting them into the LLM's embedding space. A second mechanism, CTC prompting, uses the connector's CTC output tokens as a non-autoregressive prefix prompt to the LLM, giving the autoregressive decoder a concrete guess that reduces insertion errors.

What would settle it

Run the full pipeline with the language ID labels randomly shuffled before routing and fusion; if average WER does not rise substantially, the LID-routed mLoRA is not doing the work attributed to it.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that integrating known language identification into every level of a speech LLM — encoder fusion weights, adapter selection, and the decoder — turns a generic multilingual ASR model into one that separates languages cleanly and aligns their embedding spaces. The reported numbers on the MLC-SLM development set are: baseline 20.62% average WER, the authors' re-tuned baseline 19.87%, adding dual encoders 15.95%, adding CTC loss 14.52%, adding CTC context 14.26%, and finally adding data scaling plus mLoRA 10.62%, with Qwen3-8B replacing Qwen2.5-7B in that final configuration. On the official test sets the system scores 9.60% WER for Task I and 17.49% tcpWER for Task II. The paper attributes the largest single drop to the final combination of data scaling and the LID-routed mLoRA, while noting that the CTC prompt specifically reduces insertion errors and hallucination.

Load-bearing premise

The reported gains are measured against an undertuned baseline; the paper's own minimally re-tuned baseline already closes 3.6% of the gap, so the headline 30.8% reduction depends on comparing to the original baseline rather than a tuned one.

Editorial extensions

If this is right

  • Known language ID is a usable conditioning signal for multilingual speech LLMs: routing adapters and fusion weights by LID is compatible with a shared backbone and can be trained in two stages.
  • CTC-predicted tokens can be added to an LLM-based ASR decoder as cheap context prompts, reducing insertion errors without changing the decoder's architecture.
  • A two-stage training scheme separates alignment learning (CTC) from generation learning (cross-entropy), so large amounts of transcribed speech can be used in the first stage before fine-tuning on the target corpus.
  • For diarized ASR, feeding longer, speaker-consistent segments into a strong ASR model improves downstream recognition over the baseline's fragmented segments, with gains coming from both the ASR model and a language-specialized speaker-verification model.

Reading between the lines

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

  • A clean ablation that keeps the same LLM backbone and the same 180k-hour data while toggling mLoRA on and off would isolate how much of the final 10.62% WER is due to the adapter routing itself; the published final row bundles several changes at once.
  • If the LID-routed adapter idea generalizes, it should also apply to code-switched speech, where language changes within an utterance; that would require moving from utterance-level LID to segment-level or frame-level routing.
  • The system's dependence on known LID suggests a testable follow-up: predict LID with a small classifier and route on the prediction; the drop in WER would show how much of the gain relies on perfect language identity versus approximate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This manuscript describes the TEA-ASLP submission to the MLC-SLM 2025 Challenge. For Task I (multilingual conversational ASR), the authors build on their Ideal-LLM system by combining dual Whisper/MMS encoders, a language-dependent weighted fusion connector, a CTC projector whose output is used as a prompt to the LLM decoder, and an LID-routed multilingual Mixture-of-Experts LoRA (mLoRA) adapter applied to the encoders and text decoder. Training uses roughly 180k hours of data across 11 languages in a two-stage CTC-then-CE schedule. For Task II (diarized ASR), they replace the baseline English-Chinese speaker verification model with an English-only ERes2Net-large model and concatenate adjacent same-speaker segments before ASR. The reported results are a 30.8% relative WER reduction over the challenge baseline, a final test WER of 9.60% in Task I, and a tcpWER of 17.49% in Task II, with first and second place claims in the respective tasks.

Significance. If the causal attribution in the paper were established, this would be a strong systems result: the official challenge rankings corroborate the raw performance, and the ablation table provides a mostly incremental view of the design choices (dual encoders, CTC loss, CTC context). The paper is also honest about the fact that the provided baseline is undertuned: the 'Baseline (ours)' row, which only retunes downsampling and LoRA rank, already improves the average dev WER. The main weakness is that the central named contribution, the LID-routed mLoRA, is never isolated from data scaling and a base-model swap, and the headline 30.8% figure comes from an intermediate model rather than the final submitted system. With matched ablations and corrected averages, the paper would be a useful reference for the challenge community.

major comments (3)
  1. [§3.3, Table 2] The final ablation row is labeled '+ Data Scaling & mLoRA', but the text explicitly states that this row also changes the base LLM from Qwen2.5-7B to Qwen3-8B. Relative to the '+ CTC context' row, this changes at least four variables: training data scale (1.5k to 179k hours), mLoRA, base LLM, and the training schedule/hyperparameters associated with the larger run. There is no 'data scaling only' run and no 'mLoRA only' run on a fixed base model, so the improvement from 14.26% to 10.62% cannot be attributed to mLoRA. The paper should add matched ablations, or at minimum an mLoRA on/off comparison on the same data and base model, before claiming that mLoRA is the cause of the gain. The same concern applies to the 'Baseline (ours)' row, which shows that the original baseline is undertuned; reported relative reductions should also be computed against the retuned baseline or both baselines should be reported.
  2. [Abstract and §1] The '30.8% reduction' cited in the abstract and Introduction is the dev-set result of the '+ CTC context' row, which does not include mLoRA or the 180k-hour data. The final system's dev-set reduction against the same baseline is 48.4%, and the 9.60% WER is on the test set. Pairing the 30.8% figure with the final 9.60% WER in the abstract implies that the final system achieves only the 30.8% reduction, which is misleading and inconsistent with the final row of Table 2. The abstract and Introduction should report the final system's actual relative reduction, or clearly label the 30.8% as an intermediate result.
  3. [Table 2] The 'avg' column is inconsistent with the language-wise entries in the same table. The arithmetic mean of the 11 numbers in each row is approximately 23.31, 21.98, 17.37, 16.44, 16.31, and 12.09 for the six rows, not the printed 20.62, 19.87, 15.95, 14.52, 14.26, and 10.62. Because the paper does not define the weighting used for 'avg', every relative-reduction claim derived from these averages cannot be verified. Please define the averaging scheme (for example, whether it accounts for the five English regions, language duration, or test-set composition) or correct the numbers.
minor comments (5)
  1. [§2.2, Stage 1] The sentence 'initially, both encoders are fully frozen; in the second step, the encoders are frozen, and mLoRA training is introduced' is self-contradictory. Please clarify whether the encoders are kept frozen when mLoRA is introduced, and state which parameters are trainable at each step.
  2. [§3.3, '+ Data Scaling & mLoRA' bullet] The bullet says 'The introduction of mLoRA and data scaling is reflected directly in the final results' but then acknowledges that the same row uses Qwen3-8B instead of Qwen2.5-7B. The row label and the explanatory text should be made consistent, for example by renaming the row '+ Data Scaling, mLoRA & Qwen3-8B' or by separating the changes.
  3. [Table 3] The row 'Proposed System in Task I' improves tcpWER from 60.39 to 18.56 while keeping the same MS/FA/SER values as the baseline, which is a striking result. A brief explanation of why replacing only the ASR model produces such a large tcpWER improvement would help readers interpret the pipeline contributions.
  4. [References] Reference [30] contains a formatting error ('(LREC' instead of 'LREC'), and reference [31] lists an incomplete or garbled author string ('Y. Y. D. M. S. Fujimoto'). Please correct these entries.
  5. [§2.1, Weighted fusion module] The description says that trainable weights are initialized for each language and a sigmoid is applied, but it does not state whether the fusion weights are shared across time frames or how they are applied to the sequence of hidden representations. Adding one sentence with the exact mechanism would improve reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claims are externally evaluated challenge results and controlled ablations, while the Ideal-LLM self-citation is architectural background rather than load-bearing evidence.

full rationale

The paper's load-bearing claims are (i) official challenge outcomes (9.60% WER on Task I, 17.49% tcpWER on Task II test sets, first/second places) and (ii) the incremental ablation rows in Table 2. Both are direct measurements on fixed evaluation sets or dev-set ablations, not quantities derived from a fitted parameter, a self-definitional identity, or a cited theorem. The abstract's '30.8% reduction' arithmetic equals (20.62-14.26)/20.62 from the '+ CTC context' row, not the final system row, which achieves a 48.4% reduction with bundled data scaling, mLoRA, and Qwen3-8B; Section 1 correctly attributes the 30.8% to the MLC-SLM-data-only condition, so the abstract is loose attribution, not a definitional circle. The final row bundles three changes with no isolated mLoRA ablation, leaving the named contribution without a controlled proof—a correctness/attribution gap, not circularity as defined here. The one self-citation, Ideal-LLM [17], supplies the dual-encoder/connector architecture, but the incremental gains in rows 2-4 are demonstrated by the paper's own controlled ablations, and Ideal-LLM is not invoked as evidence for the new result; no uniqueness theorem or ansatz is imported from it. The SD result similarly rests on an external speaker-verification model swap measured on the dev set. Because the central claims are self-contained against external benchmarks and no step reduces to its own input by construction, circularity is minimal; the single non-load-bearing self-citation warrants score 1 rather than 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim is an empirical benchmark result. It rests on standard pretrained models, task-supplied LID, external corpora, and a set of hand-tuned hyperparameters. There is no derivation, so the ledger mainly records the architecture choices and assumptions that would need to be controlled in a fair comparison.

free parameters (5)
  • Language-dependent fusion weights = not reported
    Trainable per-language sigmoid weights in the weighted fusion module select between Whisper and MMS features; learned via decoder and CTC loss.
  • mLoRA rank and alpha = LLM rank 64 alpha 32; encoder rank 32 alpha 16
    Chosen by hand; control adapter capacity and directly affect recognition accuracy.
  • Downsample factor = 2x
    Changed from baseline; affects time resolution of speech features fed to the LLM.
  • Data balancing weights = not reported
    Uses the strategy of Conneau et al. [35] to reweight the 180k-hour corpus across 11 languages; exact weights are not given.
  • Training schedule hyperparameters = Stage1: LR 2e-4, 2k warmup, 100k steps; Stage2: LR 5e-5, 2k warmup, 100k steps
    Fixed manually; no search or sensitivity analysis is reported.
assumptions (4)
  • domain assumption The challenge provides correct language identification for each test utterance.
    Routing of mLoRA and fusion weights depends on this; Section 2.1 says 'routed according to the specific LID' and 'Since LID is known in this task'.
  • domain assumption The official MLC-SLM evaluation protocol and test labels are reliable.
    All headline numbers are based on the challenge's dev/test sets, and the paper provides no independent evaluation.
  • domain assumption The external training corpora and pretrained models provide distributions compatible with the challenge data.
    The system transfers pretrained features and 180k hours of mostly non-conversational data to conversational test sets.
  • domain assumption OWSM-CTC filtering removes only low-quality samples and does not bias the training distribution.
    Used in Section 3.1 for YouTube data quality control, but threshold and validation are not described.
invented entities (1)
  • LID-routed multilingual MoE LoRA (mLoRA) adapter
    purpose: Routes low-rank adapters according to known language ID for dual encoders and the LLM decoder.
    The component is evaluated within this paper's challenge results, but no independent reimplementation or external validation is provided, and its specific gain is entangled with data scaling and model swap.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The TEA-ASLP System for Multilingual Conversational Speech Recognition and Speech Diarization in MLC-SLM 2025 Challenge." pith.science (2026). https://pith.science/paper/26SV2GCA

@misc{pith2026250718051,
  author       = {Pith},
  title        = {Pith review of: The TEA-ASLP System for Multilingual Conversational Speech Recognition and Speech Diarization in MLC-SLM 2025 Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/26SV2GCA}},
  note         = {Machine review of arXiv:2507.18051}
}
read the original abstract

This paper presents the TEA-ASLP's system submitted to the MLC-SLM 2025 Challenge, addressing multilingual conversational automatic speech recognition (ASR) in Task I and speech diarization ASR in Task II. For Task I, we enhance Ideal-LLM model by integrating known language identification and a multilingual MOE LoRA structure, along with using CTC-predicted tokens as prompts to improve autoregressive generation. The model is trained on approximately 180k hours of multilingual ASR data. In Task II, we replace the baseline English-Chinese speaker diarization model with a more suitable English-only version. Our approach achieves a 30.8% reduction in word error rate (WER) compared to the baseline speech language model, resulting in a final WER of 9.60% in Task I and a time-constrained minimum-permutation WER of 17.49% in Task II, earning first and second place in the respective challenge tasks.

Figures

Figures reproduced from arXiv: 2507.18051 by the authors.

Figure 1
Figure 1. The overall framework of the TEA-ASLP system. Language-adapted Connector Since the dual encoders have been trained on different language distributions, we design a connector to perform a language-dependent fusion of the dual encoders’ features, transforming them into the embedding space of the LLM. First, the speech features Fw and Fm are transformed into hidden representations by Whisper and MMS adapter, which are … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 29 canonical work pages

  1. [1]

    Recently, there has been growing interest in combining LLMs with audio encoders, en- abling the models to process and understand audio modali- ties [6, 7, 8, 9, 10, 11]

    Introduction Text-based Large Language Models (LLMs) have had a pro- found impact on the field of artificial intelligence, thanks to their advanced capabilities in understanding and generating natural language [1, 2, 3, 4, 5]. Recently, there has been growing interest in combining LLMs with audio encoders, en- abling the models to process and understand a...

    work page 2025

  2. [2]

    The TEA-ASLP System for Multilingual Conversational Speech Recognition and Speech Diarization in MLC-SLM 2025 Challenge

    Proposed System 2.1. Architecture We adopt the Ideal-LLM structure, which includes dual en- coders, a language-adapted connector, and a text decoder. To more effectively leverage the known LID information, we re- place the original LLM LoRA [19] adapter with a multilingual MoE LoRA (mLoRA) adapter, routed according to the specific LID. Additionally, we in...

    work page Pith review arXiv 2025

  3. [3]

    Datasets We use a large corpus to train our model, totaling 180k hours, as shown in Table 1

    Experiments 3.1. Datasets We use a large corpus to train our model, totaling 180k hours, as shown in Table 1. This corpus includes data from 11 lan- guages: English ( en), French ( fr), German ( de), Italian ( it), Japanese (ja), Korean (ko), Portuguese (pt), Russian (ru), Span- ish (es), Thai ( th), and Vietnamese ( vi). We apply a data bal- ancing strat...

    work page

  4. [4]

    Conclusions Our system for the MLC-SLM 2025 Challenge demonstrates significant improvements in both multilingual ASR and speech diarization tasks. By enhancing our Ideal-LLM model with lan- guage identification and a multilingual LoRA structure, and op- timizing the diarization model, we achieved notable reductions in word error rates and secured top posi...

    work page 2025

  5. [5]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023

    arXiv 2023

  6. [6]

    Introducing chatgpt,

    OpenAI, “Introducing chatgpt,” URL https://openai.com/blog/chatgpt, 2022

    work page 2022

  7. [7]

    Gpt-4 technical report,

    ——, “Gpt-4 technical report,” arXiv preprint arXiv:2308.11276, 2023

    arXiv 2023

  8. [8]

    Lan- guage models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhari- wal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Lan- guage models are few-shot learners,”Advances in neural informa- tion processing systems, vol. 33, pp. 1877–1901, 2020

    1901

Show all 42 references
  1. [9]

    Palm 2 technical report,

    R. Anil, A. M. Dai, O. Firat, M. Johnson, D. Lepikhin, A. Passos, S. Shakeri, E. Taropa, P. Bailey, Z. Chenet al., “Palm 2 technical report,” arXiv preprint arXiv:2305.10403, 2023

    2023 arXiv

  2. [10]

    Qwen2-audio technical re- port,

    Y . Chu, J. Xu, Q. Yang, 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

    2024 arXiv

  3. [11]

    Listen, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. R. Glass, “Listen, think, and understand,” in ICLR. OpenReview.net, 2024

    2024

  4. [12]

    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,” in ICLR. OpenReview.net, 2024

    2024

  5. [13]

    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

    2023 arXiv

  6. [14]

    Wavllm: Towards robust and adaptive speech large language model,

    S. Hu, L. Zhou, S. Liu, S. Chen, L. Meng, H. Hao, J. Pan, X. Liu, J. Li, S. Sivasankaran, L. Liu, and F. Wei, “Wavllm: Towards robust and adaptive speech large language model,” in Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, Novem...

    2024

  7. [15]

    Seed-asr: Understanding diverse 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, K. Gaoet al., “Seed-asr: Understanding diverse speech and contexts with llm-based speech recognition,” arXiv preprint arXiv:2407.04675, 2024

    2024 arXiv

  8. [16]

    E- chat: Emotion-sensitive spoken dialogue system with large lan- guage models,

    H. Xue, Y . Liang, B. Mu, S. Zhang, Q. Chen, and L. Xie, “E- chat: Emotion-sensitive spoken dialogue system with large lan- guage models,” in ISCSLP. IEEE, 2024

    2024

  9. [17]

    Prompt- ing large language models with speech recognition abilities,

    Y . Fathullah, C. Wu, E. Lakomkin, J. Jia, Y . Shangguan, K. Li, J. Guo, W. Xiong, J. Mahadeokar, O. Kalinli et al. , “Prompt- ing large language models with speech recognition abilities,” in ICASSP. IEEE, 2024, pp. 13 351–13 355

    2024

  10. [18]

    On decoder-only architecture for speech-to-text and large language model integration,

    J. Wu, Y . Gaur, Z. Chen, L. Zhou, Y . Zhu, T. Wang, J. Li, S. Liu, B. Ren, L. Liu, and Y . Wu, “On decoder-only architecture for speech-to-text and large language model integration,” in ASRU. IEEE, 2023, pp. 1–8

    2023

  11. [19]

    An embarrassingly simple approach 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. Zhanget al., “An embarrassingly simple approach for llm with strong asr capacity,” arXiv preprint arXiv:2402.08846 , 2024

    2024 arXiv

  12. [20]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in NeurIPS, 2017, pp. 5998–6008

    2017

  13. [21]

    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,” in ICML, vol. 202, 2023, pp. 28 492–28 518

    2023

  14. [22]

    Ideal-llm: Integrating dual encoders and language-adapted llm for multilingual speech-to-text,

    H. Xue, W. Ren, X. Geng, K. Wei, L. Li, Q. Shao, L. Yang, K. Diao, and L. Xie, “Ideal-llm: Integrating dual encoders and language-adapted llm for multilingual speech-to-text,” arXiv preprint arXiv:2409.11214, 2024

    2024 arXiv

  15. [23]

    Scaling speech technology to 1, 000+ languages,

    V . Pratap, A. Tjandra, B. Shi, P. Tomasello, A. Babu, S. Kundu, A. Elkahky, Z. Ni, A. Vyas, M. Fazel-Zarandi, A. Baevski, Y . Adi, X. Zhang, W. Hsu, A. Conneau, and M. Auli, “Scaling speech technology to 1, 000+ languages,” J. Mach. Learn. Res. , vol. 25, pp. 97:1–97:52, 2024

    2024

  16. [24]

    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 ICLR. OpenReview.net, 2022

    2022

  17. [25]

    Gigaspeech 2: An evolving, large-scale and multi-domain asr corpus for low-resource languages with au- tomated crawling, transcription and refinement,

    Y . Yang, Z. Song, J. Zhuo, M. Cui, J. Li, B. Yang, Y . Du, Z. Ma, X. Liu, Z. Wang et al., “Gigaspeech 2: An evolving, large-scale and multi-domain asr corpus for low-resource languages with au- tomated crawling, transcription and refinement,” arXiv preprint arXiv:2406.11546, 2024

    2024 arXiv

  18. [26]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv et al., “Qwen3 technical report,” arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  19. [27]

    MSR- 86K: an evolving, multilingual corpus with 86, 300 hours of tran- scribed 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 tran- scribed audio for speech recognition research,” in Interspeech. ISCA, 2024

    2024

  20. [28]

    Com- mon voice: 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 voice: A massively-multilingual speech corpus,” in LREC. European Language Resources Association, 2020, pp. 4218– 4222

    2020

  21. [29]

    MLS: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “MLS: A large-scale multilingual dataset for speech research,” in Interspeech. ISCA, 2020, pp. 2757–2761

    2020

  22. [30]

    The fisher corpus: a resource for the next generations of speech-to-text,

    C. Cieri, D. Miller, and K. Walker, “The fisher corpus: a resource for the next generations of speech-to-text,” in (LREC. European Language Resources Association (ELRA), 2004

    2004

  23. [31]

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,

    H. He, Z. Shang, C. Wang, X. Li, Y . Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shi, Y . Wang, K. Chen, P. Zhang, and Z. Wu, “Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation,” in SLT. IEEE, 2024, pp. 885–890

    2024

  24. [32]

    The average WER is reduced by 3.6% compared to the original baseline

    The training data includes only the MLC-SLM 1.5k hours dataset. The average WER is reduced by 3.6% compared to the original baseline. • + Dual Encoder: In this setup, the original Whisper En- coder is replaced by the Dual Encoder, and the proposed Language-adapted Connector is...

  25. [33]

    Opendatalab: Empowering general artificial intelligence with open datasets,

    C. He, W. Li, Z. Jin, C. Xu, B. Wang, and D. Lin, “Opendatalab: Empowering general artificial intelligence with open datasets,”

  26. [34]

    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,” in ICASSP. IEEE, 2015, pp. 5206–5210

    2015

  27. [35]

    Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,

    G. Chen, S. Chai, G. Wang, J. Du, W.-Q. Zhang, C. Weng, D. Su, D. Povey, J. Trmal, J. Zhang et al. , “Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,” arXiv preprint arXiv:2106.06909, 2021

    2021 arXiv

  28. [36]

    Owsm-ctc: An open encoder-only speech foundation model for speech recog- nition, translation, and language identification,

    Y . Peng, Y . Sudo, M. Shakeel, and S. Watanabe, “Owsm-ctc: An open encoder-only speech foundation model for speech recog- nition, translation, and language identification,” arXiv preprint arXiv:2402.12654, 2024

    2024 arXiv

  29. [37]

    Reazonspeech: A free and massive cor- pus for japanese asr,

    Y . Y . D. M. S. Fujimoto, “Reazonspeech: A free and massive cor- pus for japanese asr,” 2016

    2016

  30. [38]

    Construction of a large-scale japanese asr corpus on tv recordings,

    S. Ando and H. Fujihara, “Construction of a large-scale japanese asr corpus on tv recordings,” in ICASSP 2021-2021 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6948–6952

    2021

  31. [39]

    Golos: Russian dataset for speech research,

    N. Karpov, A. Denisenko, and F. Minkin, “Golos: Russian dataset for speech research,” arXiv preprint arXiv:2106.10161, 2021

    2021 arXiv

  32. [40]

    Ksponspeech: Korean spontaneous speech corpus for automatic speech recog- nition,

    J.-U. Bang, S. Yun, S.-H. Kim, M.-Y . Choi, M.-K. Lee, Y .-J. Kim, D.-H. Kim, J. Park, Y .-J. Lee, and S.-H. Kim, “Ksponspeech: Korean spontaneous speech corpus for automatic speech recog- nition,” Applied Sciences, vol. 10, no. 19, p. 6936, 2020

    2020

  33. [41]

    Unsupervised cross-lingual representation learning for speech recognition,

    A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, “Unsupervised cross-lingual representation learning for speech recognition,” in Interspeech. ISCA, 2021, pp. 2426–2430

    2021

  34. [2024]

    Available: https://arxiv.org/abs/2407.13773

    [Online]. Available: https://arxiv.org/abs/2407.13773

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.