Pith. sign in

REVIEW 4 major objections 5 minor 43 references

ILT-Iterative LoRA Training through Focus-Feedback-Fix for Multilingual Speech Recognition

T0 review · 4 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 By structuring LoRA fine-tuning into three stages, the paper reports consistent multilingual ASR improvements, with average word error falling to 9.20% on Whisper-large-v3 and 9.89% on Qwen2-Audio.

desk verdict A useful competition recipe paper, but the headline gains are confounded with a 40,000-hour data expansion and a 512-rank LoRA; the iterative mechanism itself is only weakly isolated. read the letter →

arxiv 2507.08477 v1 pith:PYAFJK2G submitted 2025-07-11 cs.CL

Qingliang Meng , Hao Wu , Wei Liang , Wei Xu , Qing Zhao This is my paper · ORCID
classification cs.CL
keywords multilingualASRlow-rankadaptationiterativeLoRAtrainingpseudo-labelingsupervisedfine-tuningspeechrecognitiondataaugmentationparameter-efficient
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the usual single-shot LoRA fine-tuning of speech large language models runs into an overfitting ceiling, and that this ceiling can be raised by replacing one training pass with three deliberately staged passes: Focus, Feed Back, and Fix. Each stage optimizes the same generative loss but on a different knowledge task—first fast adaptation to the target domain, then broad multilingual data expansion, then high-relevance data with pseudo-labels voted across model checkpoints. Applied to Whisper-large-v3 and Qwen2-Audio, the recipe is reported to improve word error rate on every evaluated language, bringing the average from 15.36% to 9.20% for the Whisper-based model and from 51.90% to 9.89% for the Qwen2-Audio-based model. A sympathetic reader would care because the result suggests that how LoRA updates are scheduled, not just how much data or rank is used, can move the performance of a frozen pretrained model.

What carries the argument

The load-bearing mechanism is the recursive ILT update: at round $t$, LoRA parameters $\phi_t$ are optimized on a knowledge task $k_t = (l_t, d_t)$ and merged into the current model, so the adapted weights accumulate as a weighted combination of staged updates rather than one monolithic fine-tune. In this paper the recursion runs three times, with stage-specific LoRA ranks (16, 512, 32) and target modules. Supporting machinery includes hard-voting pseudo-label generation from multiple candidate hypotheses, token-frequency analysis to pick low-frequency and difficult high-frequency text, WavLM-based cosine similarity for selecting third-party audio, and TTS-synthesized English dialect samples. All of these feed the knowledge task set $K$, which is what changes from stage to stage.

What would settle it

Train the same base model once, not in three stages, on the complete 40,000-hour augmented set with the same adapter size and the same ensemble-voted pseudo-labels, then compare development-set word error rates; if the single-pass run matches or beats 9.20% on Whisper and 9.89% on Qwen2-Audio, the three-stage iteration is not what produces the reported gain.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's terms, is that iterative LoRA training outperforms each of its component stages and every two-stage combination, and that the full Focus + Feed Back + Fix sequence yields the best results on the official development set. The paper reports consistent gains across all fifteen languages: no language regresses under the full ILT recipe. The authors attribute this to three complementary knowledge tasks: Focus Training anchors the model in the target domain without forgetting; Feed Back Training supplies 40,000 hours of data and expanded LoRA rank to fill gaps in languages such as Thai and Vietnamese; Fix Training concentrates on highly relevant selected data and ensemble-voted pseudo-labels to polish errors. Because the base models differ—one already strong at ASR, one general audio LLM with missing languages—the same three-stage loop improves both, which the paper takes as evidence that the paradigm, rather than the base model, is doing the work.

Load-bearing premise

The paper attributes the improvement to the iterative three-stage design, but every full run also receives more training data, a much larger adapter, synthesized audio, and pseudo-labels, so the specific gain from the iteration itself is never tested against a single-stage run with the same resources.

Editorial extensions

If this is right

  • If the ILT paradigm is correct, single-pass LoRA fine-tuning is not the ceiling for parameter-efficient ASR; scheduling the update as multiple staged merges can improve the final word error rate without changing the base model.
  • The three-stage recipe transfers across base models with very different pretraining coverage, which points to a general LoRA training schedule rather than a Whisper-specific or Qwen-specific trick.
  • The ablation's ordering result—Focus + Feed Back + Fix beats all pairs, while Focus + Fix hurts—implies that the middle data-expansion stage is load-bearing for the final gain, not just an optional extra.
  • The reported competition results (fourth in the ASR track, first in the speech separation and recognition track) suggest the recipe remains practical under real contest constraints of time and compute.

Reading between the lines

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

  • My inference: because the full method simultaneously grows the training set to 40,000 hours, raises LoRA rank to 512, adds synthesized and similarity-selected audio, and introduces pseudo-labels, the paper does not isolate the iterative schedule from these scaling changes; the stage structure may be a vehicle for the data increase rather than the true cause.
  • A direct experiment that would separate the two is to hold data, rank, and augmentation fixed and compare one LoRA pass with three ILT passes; if the gains vanish, the claim reduces to a data-and-capacity effect.
  • If the iteration itself is what matters, the same Focus-Feed Back-Fix ordering could be applied to other modalities and to non-ASR LoRA fine-tuning, where overfitting ceilings are also reported.
  • The pseudo-label voting component may be doing more of the Fix-stage work than the paper credits; re-running Fix with oracle labels would test whether better references rather than the stage structure drive the last gain.
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

4 major / 5 minor

Summary. The paper proposes Iterative LoRA Training (ILT), a three-stage fine-tuning recipe (Focus, Feed Back, Fix) for multilingual speech recognition, applied to Whisper-large-v3 and Qwen2-Audio. Each stage uses a different data recipe and LoRA configuration, with pseudo-label voting, TTS augmentation, and similarity-based data selection. The authors report large WER reductions on the MLC-SLM development set (from 15.36% to 9.20% for MegaWhisper and from 51.90% to 9.89% for MegaAudio) and an ablation study in Table 2. The paper also mentions the team's ranking in Interspeech 2025 MLC-SLM.

Significance. If the central claim were rigorously established, the ILT recipe would be a practically useful and potentially transferable system design for multilingual ASR fine-tuning. The paper contains a number of concrete engineering choices that are worth reporting: per-stage LoRA rank scheduling, token-frequency-based text selection, SECS-based audio selection, and majority voting for pseudo-labels. The authors also provide a partial ablation study. However, the significance is currently limited by the absence of the key control needed to attribute the gains to the iterative paradigm: a single-stage LoRA SFT run on the same union of data, with the same total training budget and comparable LoRA capacity. The reported results all use the same development set that was used for early stopping, hyperparameter selection, and checkpoint merging, which further weakens the quantitative claims. These issues are fixable in a revision, so the work is a reasonable candidate for major revision.

major comments (4)
  1. [Section 4.2, Table 2] The central attribution claim—that the three-stage iterative ILT design is responsible for the gains—is not isolated from data volume, training compute, and LoRA capacity. The full pipeline consumes the Focus data for 6 epochs, the 40,000-hour Feed Back set for 10 epochs, and the 2,791-hour Fix set for 6 epochs, with LoRA ranks of 16, 512, and 32 respectively. The 'Only Feed Back' control uses only the Feed Back recipe for 10 epochs; 'Only Focus' uses only the Focus recipe; there is no single-stage LoRA SFT trained on the union of all datasets for the full 22-epoch budget with a comparable rank. Consequently, the improvements in Table 1 over the zero-shot baselines could be explained by data scale and longer training alone, and the proposed recursive mechanism is not the only defensible explanation. This missing control is load-bearing for the paper's main claim and should be added or the claim should be weakened accordingly.
  2. [Sections 3.1 and 4.1] All headline results and ablations are evaluated on the MLC-SLM development set, but that same development set is used for early stopping, for selecting which LoRA checkpoints to merge, and for setting LoRA ranks, alphas, and data-selection thresholds. The reported WERs are therefore selection results on the tuning set. No held-out test set, challenge evaluation set, or cross-validation result is reported. The paper should either report results on a held-out set or transparently state that all numbers are development-set selections; without this, the reported differences (e.g., 9.89 vs. 10.33 in Table 2) may not reflect real generalization.
  3. [Section 2.1, Eq. (3)] Equation (3) is not mathematically well-defined. The quantity phi_t is introduced as 'coefficients capturing the effect of historical values on the present,' but it is also the object of the minimization (argmin over phi_t), and the update M_t = phi_t M_{t-1} + ... is not specified in terms of how LoRA merges are composed across stages. The statement that the iterative process 'gradually converges toward a global optimum' is asserted without assumptions, an objective, or a proof. Since this equation is the paper's only formal characterization of ILT, it should either be given a precise definition and convergence conditions or be presented explicitly as motivating intuition rather than as a mathematical model.
  4. [Section 2.3 and Section 3.2] The paper claims that the data selection and pseudo-labeling strategies 'significantly improve model performance,' but no ablation isolates these components. Table 2 only compares whole training stages; it does not measure the contribution of majority voting, the SECS audio-selection threshold, the token-frequency filtering, or the TTS augmentation. For example, one cannot tell from the reported results whether the Fix-stage gain of 0.31 percentage points comes from the selected 2,791 hours, from the pseudo-label voting, or from simply training longer on relevant data. The authors should provide component-level ablations or explicitly limit the claim to the full recipe.
minor comments (5)
  1. [Section 3.1.2] There is a notation error: the text says 'The optimal LoRA parameters are merged with M2 to produce the updated model, referred to as M2,' but the preceding stage defines the Focus-stage output as M1. The intended statement is presumably that M1 is merged with the Feed Back update to produce M2.
  2. [Section 2.1] The sentence following Eq. (3) is grammatically incomplete: 'At each iteration t, kt refers to the selected knowledge task...' should be part of a complete sentence, and the notation for phi_t is introduced inconsistently.
  3. [Abstract and Section 5] The phrase 'effectively enhancing the theoretical upper bound of model performance' claims a theoretical guarantee that the paper does not establish. The authors should replace this with a statement about empirical improvements.
  4. [Section 3.2.3] The description of the eight external datasets does not specify which portions are used in the Feed Back stage versus the Fix stage, how the 40,000-hour sample is drawn, or how the 2,791-hour relevant subset is composed from these resources. This limits reproducibility even within the challenge setting.
  5. [Equation (5) and Table 1] There are small presentation issues: 'V ote' in Eq. (5) is a typo, and Table 1's columns mix model names and system names in a way that is initially confusing; the table would be clearer if the baseline models were listed separately from the ILT systems.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline WER gains are computed on the same MLC-SLM development set used for checkpoint selection and hyperparameter tuning, so the central improvement claim is not an independent out-of-sample evaluation.

  1. fitted input called prediction [Section 3.1.1-3.1.3 (Focus/Feed Back/Fix Training) and Section 4 (Results)]
    "evaluated on the development set. The best performing LoRA parameters are merged with the base model to obtain an adapted version, referred to as M1. ... The model is trained for ten epochs and evaluated on the development set. The optimal LoRA parameters are merged with M2 ... The model is trained for six epochs and evaluated on the development set. The optimal LoRA parameters are then merged into model M2 ... All experiments are carried out on the development set of the MLC-SLM benchmark."

    Each of the three ILT stages uses the MLC-SLM development set for early stopping and checkpoint merging, and the hyperparameters and data-selection thresholds are also chosen against that same development set. Section 4 then reports Table 1's headline WERs on that same development set, presenting the values as evidence of 'consistent performance improvements across all evaluated languages' (Whisper 15.36 to 9.20; Qwen2-Audio 51.90 to 9.89). The reported result is thus not a held-out prediction: the development set is an input to model selection and is then reused as the evaluation target.

full rationale

This is an empirical challenge paper rather than a formal derivation, and there is no load-bearing self-citation chain or imported uniqueness theorem. The only substantive circularity-like defect is the reuse of the development set for both model selection and final reporting: all three stages merge 'best performing' LoRA checkpoints chosen on the MLC-SLM development set, and Table 1 reports improvements on that same set. This bias is real but not a formal tautology, because the competition rankings (4th in Track 1, 1st in Track 2) provide some external grounding, although the paper does not report test-set WERs. The absence of a same-data single-stage LoRA SFT control is a serious confounding issue but is not itself circularity. Minor self-citations, e.g., reference [30] cited for beam search, are not load-bearing. Overall score 4: the central claim still has independent content, but a key reported prediction is partly forced by the choice of evaluation target.

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

The central claim relies on several hand-set hyperparameters (LoRA ranks, epochs, data selection thresholds) and on domain assumptions about the benefit of iterative training, pseudo-label voting, and data selection. There are no invented physical or mathematical entities; the only new component is a learnable language-id embedding, which is a standard technique and is not ablated.

free parameters (4)
  • LoRA rank and alpha per stage = Stage 1: r=16, alpha=32; Stage 2: r=512, alpha=2048; Stage 3: r=32, alpha=64
    Hand-set hyperparameters for each stage, chosen without sensitivity analysis; the stage-2 rank of 512 is far above typical LoRA ranks and changes the trainable-parameter count, so it is a confound with the iterative stage's effect.
  • Epochs per stage = 6 (Focus), 10 (Feed Back), 6 (Fix)
    Chosen by monitoring dev-set WER; no early stopping criterion reported, so these are tuned values.
  • Token frequency thresholds = low < 20, high > 200, WER > 5%
    Hand-set cutoffs in Section 3.2.1 for selecting text augmentation data; the 5% WER threshold defines 'difficult cases' without stated justification.
  • SECS similarity threshold = 0.85
    Hand-set threshold in Section 3.2.2 for selecting third-party audio segments for low-resource languages.
assumptions (5)
  • ad hoc to paper Equation (3) models iterative LoRA as a recursive optimization with coefficients phi_t converging to a global optimum
    This is presented as an assumption ('The ILT approach assumes...') in Section 2.1 and is not proven; it is a notational formalism, not a theorem.
  • domain assumption Hard voting across models produces more reliable pseudo-labels
    Invoked in Section 2.3 without empirical validation of voting vs. alternative ensembling; relies on prior work on majority voting.
  • domain assumption The data selection heuristics (token frequency, WER, SECS) select data that improves recognition
    Used to construct Fix-stage data; only indirectly evaluated via the full ablation, not isolated.
  • domain assumption Qwen2-Audio has not been pretrained on Thai and Vietnamese, causing poor performance
    Claim about the pretraining data of Qwen2-Audio made in Section 3.1.2; not verifiable from the paper or the cited Qwen technical report within this text.
  • standard math LoRA decomposition and merge formula (Equation 2) are valid approximations of full fine-tuning
    Standard result from Hu et al. (2022), cited as [14]; assumed without re-derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ILT-Iterative LoRA Training through Focus-Feedback-Fix for Multilingual Speech Recognition." pith.science (2026). https://pith.science/paper/PYAFJK2G

@misc{pith2026250708477,
  author       = {Pith},
  title        = {Pith review of: ILT-Iterative LoRA Training through Focus-Feedback-Fix for Multilingual Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PYAFJK2G}},
  note         = {Machine review of arXiv:2507.08477}
}
read the original abstract

The deep integration of large language models and automatic speech recognition systems has become a promising research direction with high practical value. To address the overfitting issue commonly observed in Low-Rank Adaptation (LoRA) during the supervised fine-tuning (SFT) stage, this work proposes an innovative training paradigm Iterative LoRA Training (ILT) in combination with an Iterative Pseudo Labeling strategy, effectively enhancing the theoretical upper bound of model performance. Based on Whisper-large-v3 and Qwen2-Audio, we conduct systematic experiments using a three-stage training process: Focus Training, Feed Back Training, and Fix Training. Experimental results demonstrate the effectiveness of the proposed method. Furthermore, the MegaAIS research team applied this technique in the Interspeech 2025 Multilingual Conversational Speech Language Modeling Challenge (MLC-SLM), achieving 4th in Track 1 (Multilingual ASR Task) and 1st place in Track 2 (Speech Separation and Recognition Task), showcasing the practical feasibility and strong application potential of our approach.

Figures

Figures reproduced from arXiv: 2507.08477 by the authors.

Figure 1
Figure 1. Training paradigm and modeling. 2.1. Training Paradigm and Modeling Low-Rank Adaptation [14] has been widely applied in the field of natural language processing. By freezing the parameters of the original large language model and updating only the low￾rank incremental weight matrices, LoRA enables efficient adap￾tation of the model to downstream tasks. The corresponding formulation is provided in Equation (1). Durin… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 20 canonical work pages

  1. [1]

    Introduction Multilingual automatic speech recognition (ASR) has long been a central focus in the field of speech recognition[1, 2, 3, 4]. This area faces two core challenges: first, how to build a unified model capable of accurately recognizing speech across multiple languages and dialects; second, the scarcity of data resources for low-resource language...

    work page Pith review arXiv 2025

  2. [2]

    A detailed overview of the workflow is presented in Figure 1

    Methodology In this section, we elaborate on the model training procedure and the adopted modeling methodology. A detailed overview of the workflow is presented in Figure 1. Figure 1: Training paradigm and modeling. 2.1. Training Paradigm and Modeling Low-Rank Adaptation [14] has been widely applied in the field of natural language processing. By freezing...

    work page

  3. [3]

    Three-Stage Iterative Training Strategy Considering the time constraints of the competition, we em- ployed only three iterations of Iterative LoRA Training (3)

    Experiment 3.1. Three-Stage Iterative Training Strategy Considering the time constraints of the competition, we em- ployed only three iterations of Iterative LoRA Training (3). Specifically, the training process consisted of three consecutive stages: Focus Training, Feed Back Training, and Fix Training. 3.1.1. Focus Training Speech large language models a...

    work page

  4. [4]

    All experiments are carried out on the development set of the MLC- SLM benchmark 2

    Results This section evaluates the effectiveness of the proposed ap- proach and conducts ablation studies on each training stage. All experiments are carried out on the development set of the MLC- SLM benchmark 2. To distinguish the ILT-adapted models from their original versions, we refer to the Whisper-large-v3 model after ILT training as MegaWhisper, a...

    work page

  5. [5]

    In the MLC-SLM challenge, we design a three-stage training paradigm consisting of Focus, Feed Back, and Fix Training

    Conclusions This study proposes an iterative fine-tuning framework based on Low Rank Adaptation, aiming to enhance model perfor- mance through a knowledge task oriented adjustment mecha- nism. In the MLC-SLM challenge, we design a three-stage training paradigm consisting of Focus, Feed Back, and Fix Training. Each stage is carefully aligned with a specifi...

    work page

  6. [6]

    Massively multilingual asr on 70 lan- guages: Tokenization, architecture, and generalization capabili- ties,

    A. Tjandra, N. Singhal, D. Zhang, O. Kalinli, A. Mohamed, D. Le, and M. L. Seltzer, “Massively multilingual asr on 70 lan- guages: Tokenization, architecture, and generalization capabili- ties,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    work page 2023

  7. [7]

    Massively multilingual asr: 50 languages, 1 model, 1 billion parameters,

    V . Pratap, A. Sriram, P. Tomasello, A. Hannun, V . Liptchin- sky, G. Synnaeve, and R. Collobert, “Massively multilingual asr: 50 languages, 1 model, 1 billion parameters,” arXiv preprint arXiv:2007.03001, 2020

    arXiv 2007

  8. [8]

    Scaling end-to- end models for large-scale multilingual asr,

    B. Li, R. Pang, T. N. Sainath, A. Gulati, Y . Zhang, J. Qin, P. Haghani, W. R. Huang, M. Ma, and J. Bai, “Scaling end-to- end models for large-scale multilingual asr,” in 2021 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 1011–1018

    work page 2021

Show all 43 references
  1. [9]

    Multi- ple softmax architecture for streaming multilingual end-to-end asr systems

    V . Joshi, A. Das, E. Sun, R. R. Mehta, J. Li, and Y . Gong, “Multi- ple softmax architecture for streaming multilingual end-to-end asr systems.” in Interspeech, 2021, pp. 1767–1771

    2021

  2. [10]

    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,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

    2023

  3. [11]

    Camel: Cross-attention enhanced mixture-of-experts and lan- guage bias for code-switching speech recognition,

    H. Wang, X. Wan, N. Zheng, K. Liu, H. Zhou, G. Li, and L. Xie, “Camel: Cross-attention enhanced mixture-of-experts and lan- guage bias for code-switching speech recognition,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)....

    2025

  4. [12]

    Owsm v3. 1: Better and faster open whisper-style speech models based on e-branchformer,

    Y . Peng, J. Tian, W. Chen, S. Arora, B. Yan, Y . Sudo, M. Shakeel, K. Choi, J. Shi, X. Chang et al., “Owsm v3. 1: Better and faster open whisper-style speech models based on e-branchformer,” arXiv preprint arXiv:2401.16658, 2024

    2024 arXiv

  5. [13]

    Dolphin: A large-scale automatic speech recognition model for eastern languages,

    Y . Meng, J. Li, G. Lin, Y . Pu, G. Wang, H. Du, Z. Shao, Y . Huang, K. Li, and W.-Q. Zhang, “Dolphin: A large-scale automatic speech recognition model for eastern languages,” arXiv preprint arXiv:2503.20212, 2025

    2025 arXiv

  6. [14]

    Llama: Open and efficient foundation language models,

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

    2023 arXiv

  7. [15]

    Qwen technical report,

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huanget al., “Qwen technical report,” arXiv preprint arXiv:2309.16609, 2023

    2023 arXiv

  8. [16]

    Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,” arXiv preprint arXiv:2305.11000, 2023

    2023 arXiv

  9. [17]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Linet al., “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    2024 arXiv

  10. [18]

    Mini-omni: Language models can hear, talk while thinking in streaming,

    Z. Xie and C. Wu, “Mini-omni: Language models can hear, talk while thinking in streaming,” arXiv preprint arXiv:2408.16725 , 2024

    2024 arXiv

  11. [19]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al. , “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  12. [20]

    Lora-whisper: Parameter-efficient and extensible multilingual asr,

    Z. Song, J. Zhuo, Y . Yang, Z. Ma, S. Zhang, and X. Chen, “Lora-whisper: Parameter-efficient and extensible multilingual asr,” arXiv preprint arXiv:2406.06619, 2024

    2024 arXiv

  13. [21]

    Towards rehearsal-free multilingual asr: A lora-based case study on whisper,

    T. Xu, K. Huang, P. Guo, Y . Zhou, L. Huang, H. Xue, and L. Xie, “Towards rehearsal-free multilingual asr: A lora-based case study on whisper,” arXiv preprint arXiv:2408.10680, 2024

    2024 arXiv

  14. [22]

    Hdmole: Mixture of lora experts with hierarchical routing and dynamic thresholds for fine-tuning llm-based asr models,

    B. Mu, K. Wei, Q. Shao, Y . Xu, and L. Xie, “Hdmole: Mixture of lora experts with hierarchical routing and dynamic thresholds for fine-tuning llm-based asr models,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2...

    2025

  15. [23]

    Iterative pseudo-labeling for speech recognition,

    Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert, “Iterative pseudo-labeling for speech recognition,” arXiv preprint arXiv:2005.09267, 2020

    2005 arXiv

  16. [24]

    Improved noisy iterative pseudo- labeling for semi-supervised speech recognition,

    T. Li, Q. Meng, and Y . Sun, “Improved noisy iterative pseudo- labeling for semi-supervised speech recognition,” in 2022 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2023, pp. 167–173

    2022

  17. [25]

    Improved noisy student training for automatic speech recognition,

    D. S. Park, Y . Zhang, Y . Jia, W. Han, C.-C. Chiu, B. Li, Y . Wu, and Q. V . Le, “Improved noisy student training for automatic speech recognition,” arXiv preprint arXiv:2005.09629, 2020

    2005 arXiv

  18. [26]

    Semi-supervised learning with data augmentation for end-to-end asr,

    F. Weninger, F. Mana, R. Gemello, J. Andr ´es-Ferrer, and P. Zhan, “Semi-supervised learning with data augmentation for end-to-end asr,” arXiv preprint arXiv:2007.13876, 2020

    2007 arXiv

  19. [27]

    Chinchunmei at wassa 2024 empathy and personality shared task: Boosting llm’s predic- tion with role-play augmentation and contrastive reasoning cali- bration,

    T. Li, N. Rusnachenko, and H. Liang, “Chinchunmei at wassa 2024 empathy and personality shared task: Boosting llm’s predic- tion with role-play augmentation and contrastive reasoning cali- bration,” in Proceedings of the 14th Workshop on Computational Approaches to Subjectivit...

    2024

  20. [28]

    Periodiclora: Break- ing the low-rank bottleneck in lora optimization,

    X. Meng, D. Dai, W. Luo, Z. Yang, S. Wu, X. Wang, P. Wang, Q. Dong, L. Chen, and Z. Sui, “Periodiclora: Break- ing the low-rank bottleneck in lora optimization,” arXiv preprint arXiv:2402.16141, 2024

    2024 arXiv

  21. [29]

    Deliberate practice and acquisition of ex- pert performance: a general overview,

    K. Anders Ericsson, “Deliberate practice and acquisition of ex- pert performance: a general overview,” Academic emergency medicine, vol. 15, no. 11, pp. 988–994, 2008

    2008

  22. [30]

    V oting: A machine learning approach,

    D. Burka, C. Puppe, L. Szepesv ´ary, and A. Tasn ´adi, “V oting: A machine learning approach,” European Journal of Operational Research, vol. 299, no. 3, pp. 1003–1017, 2022

    2022

  23. [31]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  24. [32]

    Spark-tts: An efficient llm- based text-to-speech model with single-stream decoupled speech tokens,

    X. Wang, M. Jiang, Z. Ma, Z. Zhang, S. Liu, L. Li, Z. Liang, Q. Zheng, R. Wang, X. Feng et al., “Spark-tts: An efficient llm- based text-to-speech model with single-stream decoupled speech tokens,” arXiv preprint arXiv:2503.01710, 2025

    2025 arXiv

  25. [33]

    Wavlm: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  26. [34]

    Sc-glowtts: An efficient zero-shot multi-speaker text-to-speech model. arxiv 2021,

    E. Casanova, C. Shulby, E. G ¨olge, N. M¨uller, F. de Oliveira, A. Ju- nior, A. da Silva Soares, S. Aluisio, and M. Ponti, “Sc-glowtts: An efficient zero-shot multi-speaker text-to-speech model. arxiv 2021,” arXiv preprint arXiv:2104.05557, 2021

    2021 arXiv

  27. [35]

    Mtlm: an innova- tive language model training paradigm for asr,

    Q. Meng, P. Ren, T. Li, and C. Dai, “Mtlm: an innova- tive language model training paradigm for asr,” arXiv preprint arXiv:2502.10058, 2025

    2025 arXiv

  28. [36]

    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,” arXiv preprint arXiv:2012.03411, 2020

    2012 arXiv

  29. [37]

    The multilingual tedx corpus for speech recognition and translation,

    E. Salesky, M. Wiesner, J. Bremerman, R. Cattoni, M. Negri, M. Turchi, D. W. Oard, and M. Post, “The multilingual tedx corpus for speech recognition and translation,” arXiv preprint arXiv:2102.01757, 2021

    2021 arXiv

  30. [38]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,”arXiv preprint arXiv:1912.06670, 2019

    1912 arXiv

  31. [39]

    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

  32. [40]

    Maxseats meeting valid dataset,

    Maxseats, “Maxseats meeting valid dataset,” https://huggingface.co/datasets/maxseats/meeting valid

  33. [41]

    Yodas: Youtube-oriented dataset for audio and speech,

    X. Li, S. Takamichi, T. Saeki, W. Chen, S. Shiota, and S. Watan- abe, “Yodas: Youtube-oriented dataset for audio and speech,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  34. [42]

    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,” null, 2016

    2016

  35. [43]

    Santa Barbara corpus of spoken Amer- ican English, Parts 1–4,

    J. W. Du Bois, W. L. Chafe, C. Meyer, S. A. Thompson, R. Engle- bretson, and N. Martey, “Santa Barbara corpus of spoken Amer- ican English, Parts 1–4,” Linguistic Data Consortium, Philadel- phia, 2000–2005

    2000

Pith tools

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