Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Efficient Multilingual ASR Finetuning via LoRA Language Experts

T0 review · 3 major / 6 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 Per-language LoRA experts, fused into a routing mixture or distilled into one student, improve Whisper-based multilingual ASR by about 10% relative WER with language ID and 15% without.

desk verdict The LoRA MoLE recipe is genuinely useful, but the headline KD gains are confounded by the averaged-expert initialization, so the central attribution needs a control before I'd trust it. read the letter →

arxiv 2506.21555 v1 pith:Q4VRQQIB submitted 2025-06-11 cs.CL cs.SDeess.AS

Jiahong Li , Yiwen Shao , Jianheng Zhuo , Chenda Li , Liliang Tang , Dong Yu , Yanmin Qian This is my paper · ORCID
classification cs.CLcs.SDeess.AS
keywords multilingualspeechrecognitionLoRAlanguageexpertsknowledgedistillationmixtureofWhisperparameter-efficientfinetuningexpansion
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

Multilingual speech recognition suffers from the curse of multilinguality: training one model on many languages makes languages interfere, hurting accuracy. The paper proposes to prepare a separate low-rank LoRA adapter for each target language, then combine those adapters into one system in either of two ways: merge their shallow layers into a routing mixture (LoRA MoLE) for language-agnostic use, or distill all the experts layer-wise into a single larger-rank LoRA student. On a Whisper-medium backbone and eight languages, the distilled student reduces average word error rate by about 10% relative to a standard multilingual LoRA finetune when the language is provided, and by about 15% when the language must be inferred. The practical appeal is efficiency and extensibility: the per-language experts are trained and frozen in advance, and the final deployed model is a single set of LoRA parameters rather than an ever-growing collection of adapters.

What carries the argument

The load-bearing objects are LoRA language experts: for each target language, a set of low-rank matrices $\Delta W = BA$ attached to the frozen Whisper weights, trained monolingually to convergence. Two recombination mechanisms carry the argument. The first is LoRA MoLE, in which the first $L_1$ encoder layers' expert matrices are summed with a learned softmax weighting to form shared multilingual layers, and a small MLP predicts language routes; this turns per-language adapters into one language-agnostic model with few trainable parameters. The second is layer-wise knowledge distillation, in which a single larger-rank LoRA student is initialized from the average of the experts and trained to match each expert's hidden states by cosine similarity while the teacher and student outputs are interpolated with probability 0.5 before the next layer, plus a Jensen-divergence loss on the final ASR logits. The interpolation step is what lets the student follow expert representations without cascading mismatch, and the residual, additive form of LoRA is what makes the stated assumption of minor differences between expert representations plausible.

What would settle it

Compare the average cosine similarity between the hidden states of the eight per-language LoRA experts at corresponding layers on a shared set of utterances; if the similarity is low or varies strongly across language pairs, the assumption of minor differences is false and the interpolation-based KD would be expected to degrade. Alternatively, training the same recipe on languages from distant families and observing the 15% language-agnostic gain shrink would falsify the general claim.

Watch

Extended reading notes

Core claim

The central claim is that the interference problem of multilingual ASR can be reduced by decoupling language-specific knowledge into separately trained LoRA experts and then recombining that knowledge at a lower cost than direct multilingual finetuning. With eight experts on top of Whisper-medium, the paper reports average WER of 8.74% for the distilled student when language ID is given, versus 9.72% for the multilingual LoRA baseline (about 10% relative improvement), and 9.70% versus 11.51% without language ID (about 15% relative improvement). The language-agnostic version, LoRA MoLE, merges the first 20 encoder layers of the experts using learned softmax weights and keeps a small MLP for language routing, reaching its best performance with only about 1M trainable parameters. The paper also shows that the distilled student's hidden representations separate languages more cleanly than those of a directly finetuned LoRA model, which it interprets as successful layer-wise alignment with the experts.

Load-bearing premise

The distillation recipe depends on the assumption, stated in Section 3.4, that the layer-wise hidden representations produced by the different LoRA experts differ only slightly, so a single shared student with larger rank can match all of them; if that near-similarity fails for other languages or backbones, the language-agnostic gain is not guaranteed.

Editorial extensions

If this is right

  • A practical recipe emerges: prepare one LoRA expert per language, freeze it, and train either a fused router or a distilled student; both beat a multilingual LoRA model trained directly on the mixed data.
  • The distilled student gives most of the per-language experts' accuracy in the language-aware setting and closes much of the gap in the language-agnostic setting, so a single adapted model can serve both use cases.
  • Language expansion becomes additive: a newly added language needs only one new expert, and the existing frozen experts and the Whisper backbone do not have to be retrained.
  • Because LoRA MoLE updates only learnable combination weights and a routing MLP, a language-agnostic multilingual system can be built with far fewer trainable parameters than full finetuning.

Reading between the lines

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

  • If the near-similarity of expert representations holds beyond this eight-language set, the same distillation recipe could be applied to any frozen multilingual encoder, not only Whisper, since the method only touches residual adapters and layer-wise losses.
  • The paper tests one expansion language, Cantonese; a natural stress test is to add many low-resource or typologically distant languages and watch whether the 15% language-agnostic gain decays, which would indicate where the hidden-representation assumption breaks.
  • The student's rank (256) is four times each expert's rank (64), so the distillation may effectively be learning a shared subspace that covers all experts; comparing student ranks and expert-count/rank ratios could reveal a scaling law for the 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

3 major / 6 minor

Summary. The paper proposes an efficient multilingual ASR fine-tuning framework built on Whisper-medium, with three components: per-language LoRA experts, a LoRA mixture of language experts (MoLE) that merges shallow encoder layers via learned weights and an MLP router, and a single multilingual LoRA student distilled layer-wise from the experts. Experiments on eight languages (seven base languages plus Cantonese as language expansion) report approximately 10% and 15% relative WER gains over a standard multilingual LoRA baseline in language-aware and language-agnostic scenarios, respectively. Ablations vary the number of merged MoLE layers, the student rank, and the distillation signal (layer-wise versus logits-only). The central claim is that prepared LoRA experts, combined by fusion or distillation, beat standard fine-tuning in both efficiency and recognition performance.

Significance. If the causal attribution holds, the framework is practically interesting: it exploits already-trained monolingual LoRA experts as teachers, reduces trainable parameters in the MoLE variant to 1M, and achieves better WER than a standard multilingual LoRA baseline on a diverse language set. The paper is internally consistent and provides a clear experimental layout, including ablations and a t-SNE analysis. However, the empirical evidence is weakened by a missing control for the averaged-expert initialization, the absence of error bars or validation-selection details, and an unquantified load-bearing assumption about the similarity of LoRA expert representations. The contribution is a promising recipe, but the reported gain magnitudes are not yet causally established.

major comments (3)
  1. [§3.4, Table 1, Table 2] The claimed relative gains (10% language-aware and 15% language-agnostic, stated in the Abstract and Table 1) cannot currently be attributed to layer-wise knowledge distillation, because the LoRA-KD student is initialized from the average value of LoRA experts (Section 3.4) while the Multi. LoRA baseline is not given this initialization. Averaging per-language LoRA weights is itself a model-merging operation and may provide a better starting point than the baseline before any KD loss contributes. Table 2's ablations (rank 128, logits-only) all retain the averaged-init/training recipe, so they do not isolate the distillation signal. Please add at least two controls: (a) Multi. LoRA trained from the averaged-expert initialization without any layer-wise KD, and (b) LoRA-KD trained from a random or standard initialization with the same KD loss, reporting both language-aware and language-agnostic WER. Without these controls, the central attribution of the improvements to the KD mechanism is unestablished.
  2. [§4.1–§4.2, Figure 4] The selection of the Curated Fully-FT language sampling weights ('as concluded from experiments') and the merged-layer count L1=20 in Figure 4 appears to be based on test-set metrics, and no validation split or run-to-run variance is reported anywhere. This risks overfitting the comparison to the chosen test sets and makes it impossible to tell whether reported differences (e.g., 8.74 vs. 9.72 for language-aware average WER in Table 1) are within noise. Please specify the validation protocol and report mean and standard deviation over at least three training seeds for the main systems.
  3. [§3.4, Eq. (11), Figure 3] The method's central assumption that 'latent representations from different LoRA experts exhibit only minor differences' is unquantified. The t-SNE visualization in Figure 3 is post-hoc and qualitative, and no language-identification accuracy is reported for the distilled student despite the paper attributing part of the language-agnostic gain to better language separation. Provide a quantitative measure of inter-expert representation similarity (e.g., CKA or cosine distance of layer outputs before distillation) and report LID accuracy; otherwise, the transferability of the KD gains to more diverse language sets or other backbones remains unsupported.
minor comments (6)
  1. [§3.3, §3.4] The symbol α is used both for the MoLE combination weights in Eq. (5) and for the KD balancing weight in Eq. (12); please rename one to avoid confusion.
  2. [§3.2, Eq. (4)] The notation for LoRA experts is inconsistent: Section 3.2 defines A_i,j and B_i,j, while Eq. (4) writes the update as B_i A_i; please unify the notation.
  3. [§4.1] The sentence 'rank of 64 for monolingual LoRA experts and 256 for multilingual LoRA baseline and teacher model' is ambiguous, since the LoRA experts themselves are the teachers; please clarify whether the teacher in the KD setup is a rank-256 model or the ensemble of rank-64 experts.
  4. [Table 1] The entry '8.888.32' for LoRA MoLE in the language-agnostic column appears to be a typo, likely meaning '8.88, 8.32'; please correct it.
  5. [§3.4, §4.2] The language-agnostic inference procedure for the LoRA-KD student is not described; please specify how the absence of LID is simulated at test time (e.g., whether the initial language token is omitted, predicted greedily, or handled by the model).
  6. [§4.4] The limitation section lists only the growth in parameter storage; it should also mention the missing initialization control and the unverified similarity assumption, as these affect the interpretation of the reported gains.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical benchmark against external test sets, with no derivation that reduces to its own inputs.

full rationale

The paper reports an empirical finetuning recipe rather than a formal derivation. Its central claims are average WER comparisons in Table 1 against internally trained baselines on external test sets (AISHELL, LibriSpeech, CommonVoice). The LoRA-KD student is initialized from the average of the LoRA experts and then distilled, but this is part of the proposed training procedure, not a fitted parameter disguised as a prediction. The only self-citation is the prior LoRA-Whisper work [21] in related work, which is not load-bearing for the reported gains. Section 3.4's assumption that expert latent representations differ only slightly is an explicit premise of the method, not a conclusion drawn from the experimental results. The absence of an ablation isolating the averaged initialization from the distillation loss is a real attribution limitation but not definitional circularity; it does not make the reported benchmark result equivalent to its inputs by construction.

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

The framework introduces no new physical or architectural entities beyond LoRA updates. Its empirical claims rest on tuned hyperparameters (merged-layer count, sampling weights, interpolation probability) and on the unverified assumption about representation similarity across experts.

free parameters (4)
  • LoRA MoLE merged-layer count L1 = 20
    Chosen via ablation on the test benchmarks (Sec 4.3, Figure 4); affects routing accuracy and ASR WER.
  • Curated Fully-FT language sampling weights = {1.0,1.0,0.25,0.25,0.25,0.25,0.25,0.5}
    Described as 'concluded from experiments' (Sec 4.2), i.e., tuned to improve test WER; used as a comparison baseline.
  • Layer-wise KD interpolation probability = 0.5
    Set by hand (Sec 3.4); no sensitivity analysis reported.
  • KD balancing weight alpha = not reported
    In Eq. 12, alpha is not given numerically; ablation does not sweep it.
assumptions (4)
  • domain assumption Whisper-medium, pretrained on large-scale multilingual weakly supervised data, provides a reasonable base for finetuning to all eight target languages.
    Stated in Sec 2.1; the method's benefits are demonstrated only on this backbone.
  • domain assumption The internal multilingual training corpus is representative of the public test benchmarks (AISHELL-1, LibriSpeech, CommonVoice).
    Training and test domains are not matched explicitly; internal data details are not provided.
  • ad hoc to paper Latent representations from different LoRA language experts exhibit only minor differences.
    Explicit assumption in Sec 3.4 that justifies single-student layer-wise KD; not verified experimentally beyond t-SNE visualization.
  • standard math Standard matrix multiplication and Transformer layer operations as in Whisper.
    Used implicitly throughout; not questionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Multilingual ASR Finetuning via LoRA Language Experts." pith.science (2026). https://pith.science/paper/Q4VRQQIB

@misc{pith2026250621555,
  author       = {Pith},
  title        = {Pith review of: Efficient Multilingual ASR Finetuning via LoRA Language Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q4VRQQIB}},
  note         = {Machine review of arXiv:2506.21555}
}
read the original abstract

Recent advancements in deep learning have significantly enhanced multilingual automatic speech recognition (ASR) due to the development of advanced model architectures and available large-scale multilingual datasets. Despite that, multilingual ASR still suffers from the curse of multilinguality in that different languages tend to interfere with each other, making it difficult for the ASR model to identify multiple languages effectively while sharing model capacity across them. This paper proposes an efficient finetuning framework for customized multilingual ASR via prepared LoRA language experts based on Whisper. Through LoRA expert fusion or knowledge distillation, our approach achieves better recognition performance on target languages than standard fine-tuning methods. Experimental results demonstrate that the proposed models yield approximately 10\% and 15\% relative performance gains in language-aware and language-agnostic scenarios, respectively.

Figures

Figures reproduced from arXiv: 2506.21555 by the authors.

Figure 1
Figure 1. Paradigm of the MoLE encoder with combined LoRA language experts. is finetuned on sufficient monolingual ASR data until conver￾gence. LoRA language experts perform better than a single multilingual model due to the curse of multilinguality, where different languages can interfere with each other, causing per￾formance degradation. Information from various languages in the database can be effectively stored within the… view at source ↗
Figure 3
Figure 3. Visualization of the encoded hidden representations by t-SNE for different models. 4. Experiments 4.1. Setup The backbone model is Whisper-medium which is well￾pretrained to produce reasonable transcriptions for a wide range of languages. We select Zh (Chinese), En (English), Ko (Ko￾rean), Ja (Japanese), Ru (Russian), Vi (Vietnamese), and Id (In￾donesian) as base languages and Yue (Cantonese) as language expansion. … view at source ↗
Figure 4
Figure 4. Variation of language routing accuracy and recog￾nition WER regarding the number of merged layers in LoRA MoLE. Whisper encoder has 24 layers in total [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 34 canonical work pages

  1. [1]

    With the advent of deep learning techniques, ASR has achieved remarkable suc- cess in various real-world scenarios[1, 2, 3, 4, 5]

    Introduction As a crucial component of human-machine interaction, auto- matic speech recognition (ASR) focuses on transcribing speech signals into their corresponding written text. With the advent of deep learning techniques, ASR has achieved remarkable suc- cess in various real-world scenarios[1, 2, 3, 4, 5]. Meanwhile, multilingual ASR, which necessitat...

    work page

  2. [2]

    Efficient Multilingual ASR Finetuning via LoRA Language Experts

    Background 2.1. Whisper In this work, we adopt Whisper-medium[9] as the multilingual foundation model for convenience, yet other foundation models should have similar behaviors. Whisper[9] is a typical encoder- decoder Transformer model designed for multiple speech pro- cessing tasks, including multilingual speech recognition, speech translation, language...

    work page Pith review arXiv 2025

  3. [3]

    Methods In this paper, we propose an efficient finetuning framework that leverages monolingual LoRA language experts to facilitate the finetuning process of customized multilingual ASR models. 3.1. Task description Customized multilingual ASR models are designed to target a specific range of languages, encompassing both the base lan- guages included durin...

    work page

  4. [4]

    Setup The backbone model is Whisper-medium which is well- pretrained to produce reasonable transcriptions for a wide range of languages

    Experiments 4.1. Setup The backbone model is Whisper-medium which is well- pretrained to produce reasonable transcriptions for a wide range of languages. We select Zh (Chinese), En (English), Ko (Ko- rean), Ja (Japanese), Ru (Russian), Vi (Vietnamese), and Id (In- donesian) as base languages and Yue (Cantonese) as language expansion. Experiments are condu...

    work page

  5. [5]

    Conclusion In this paper, we introduce an efficient finetuning framework for customized multilingual ASR based on foundation models such as Whisper. By leveraging language-specific LoRA language experts, we can train a language-agnostic LoRA MoLE model at minimal cost, and achieve superior performance on both testing scenarios with a multilingual LoRA stu...

    work page

  6. [6]

    Acknowledgement This work was supported in part by China NSFC projects under Grants 62122050 and 62071288, in part by Shanghai Munici- pal Science and Technology Commission Project under Grant 2021SHZDZX0102, and in part by Tencent AI Lab Rhino-Bird Focused Research Program

    work page

  7. [7]

    Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,” inProc. ICML, 2006, pp. 369–376

    work page 2006

  8. [8]

    Speech recognition with deep recurrent neural networks,

    A. Graves, A.-r. Mohamed, and G. Hinton, “Speech recognition with deep recurrent neural networks,” inProc. IEEE ICASSP, 2013, pp. 6645–6649

    work page 2013

Show all 38 references
  1. [9]

    Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

    W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” inProc. IEEE ICASSP, 2016, pp. 4960– 4964

    2016

  2. [10]

    Joint CTC-attention based end-to-end speech recognition using multi-task learning,

    S. Kim, T. Hori, and S. Watanabe, “Joint CTC-attention based end-to-end speech recognition using multi-task learning,” inProc. IEEE ICASSP. IEEE, 2017, pp. 4835–4839

    2017

  3. [11]

    Recent advances in end-to-end automatic speech recognition,

    J. Liet al., “Recent advances in end-to-end automatic speech recognition,”APSIPA Trans. on Signal and Information Process- ing, vol. 11, no. 1, 2022

    2022

  4. [12]

    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 Proc. ISCA Interspeech, 2020, pp. 2757–2761

    2020

  5. [13]

    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. IEEE SLT, 2023, pp. 798–805

    2023

  6. [14]

    Common V oice: 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 V oice: A massively-multilingual speech corpus,” in Proc. ACL, 2020, pp. 4218–4222

    2020

  7. [15]

    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,” inProc. ICML, 2023, pp. 28 492–28 518

    2023

  8. [16]

    Google USM: Scaling auto- matic speech recognition beyond 100 languages,

    Y . Zhang, W. Han, J. Qin, Y . Wang, A. Bapna, Z. Chen, N. Chen, B. Li, V . Axelrod, G. Wanget al., “Google USM: Scaling auto- matic speech recognition beyond 100 languages,”arXiv preprint arXiv:2303.01037, 2023

    2023 arXiv

  9. [17]

    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-Zarandiet al., “Scaling speech technology to 1,000+ languages,”Journal of Machine Learning Research, vol. 25, no. 97, pp. 1–52, 2024

    2024

  10. [18]

    Mixture of informed experts for multilingual speech recognition,

    N. Gaur, B. Farris, P. Haghani, I. Leal, P. J. Moreno, M. Prasad, B. Ramabhadran, and Y . Zhu, “Mixture of informed experts for multilingual speech recognition,” inProc. IEEE ICASSP, 2021, pp. 6234–6238

    2021

  11. [19]

    A configurable multilingual model is all you need to recognize all languages,

    L. Zhou, J. Li, E. Sun, and S. Liu, “A configurable multilingual model is all you need to recognize all languages,” inProc. IEEE ICASSP, 2022, pp. 6422–6426

    2022

  12. [20]

    Language-routing mixture of experts for multilingual and code-switching speech recognition,

    W. Wang, G. Ma, Y . Li, and B. Du, “Language-routing mixture of experts for multilingual and code-switching speech recognition,” inProc. ISCA Interspeech, 2023, pp. 1389–1393

    2023

  13. [21]

    Enhancing mul- tilingual speech recognition through language prompt tuning and frame-level language adapter,

    S. Li, Y . You, X. Wang, K. Ding, and G. Wan, “Enhancing mul- tilingual speech recognition through language prompt tuning and frame-level language adapter,” inProc. IEEE ICASSP, 2024, pp. 10 941–10 945

    2024

  14. [22]

    Large-scale multilin- gual speech recognition with a streaming end-to-end model,

    A. Kannan, A. Datta, T. N. Sainath, E. Weinstein, B. Ramabhad- ran, Y . Wu, A. Bapna, Z. Chen, and S. Lee, “Large-scale multilin- gual speech recognition with a streaming end-to-end model,” in Proc. ISCA Interspeech, 2019, pp. 2130–2134

    2019

  15. [23]

    Weighted cross-entropy for low-resource languages in multilingual speech recognition,

    A. Pi ˜neiro-Mart´ın, C. Garc´ıa-Mateo, L. Docio-Fernandez, M. del Carmen L´opez-P´erez, and G. Rehm, “Weighted cross-entropy for low-resource languages in multilingual speech recognition,” in Proc. ISCA Interspeech, 2024, pp. 1235–1239

    2024

  16. [24]

    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,” inProc. ISCA Interspeech, 2024, pp. 2534– 2538

    2024

  17. [25]

    Continual learning op- timizations for auto-regressive decoder of multilingual ASR sys- tems,

    C. Y . Kwok, J. Q. Yip, and E. S. Chng, “Continual learning op- timizations for auto-regressive decoder of multilingual ASR sys- tems,” inProc. ISCA Interspeech, 2024, pp. 1225–1229

    2024

  18. [26]

    Extending multilingual ASR to new languages using supplementary encoder and decoder components,

    Y . Khassanov, Z. Chen, T. Chen, T. Y . Chong, W. Li, L. Lu, and Z. Ma, “Extending multilingual ASR to new languages using supplementary encoder and decoder components,” inProc. IEEE ICASSP, 2024, pp. 10 586–10 590

    2024

  19. [27]

    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,” inProc. ISCA Interspeech, 2024, pp. 3934–3938

    2024

  20. [28]

    A parameter- efficient language extension framework for multilingual ASR,

    W. Liu, J. Hou, D. Yang, M. Cao, and T. Lee, “A parameter- efficient language extension framework for multilingual ASR,” in Proc. ISCA Interspeech, 2024, pp. 3929–3933

    2024

  21. [29]

    Improving multi- lingual ASR in the wild using simple N-best re-ranking,

    B. Yan, V . Pratap, S. Watanabe, and M. Auli, “Improving multi- lingual ASR in the wild using simple N-best re-ranking,”arXiv preprint arXiv:2409.18428, 2024

    2024 arXiv

  22. [30]

    LoRA: Low-rank adaptation of large lan- guage models,

    E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large lan- guage models,” inProc. ICLR, 2022

    2022

  23. [31]

    Knowl- edge distillation for mixture of experts models in speech recogni- tion,

    F. C. Salinas, K. Kumatani, R. Gmyr, L. Liu, and Y . Shi, “Knowl- edge distillation for mixture of experts models in speech recogni- tion,” Microsoft Tech. Report, MSR-TR-2022-6, May 2022, Tech. Rep., 2022

    2022

  24. [32]

    Multilingual Distilwhisper: Efficient distillation of multi-task speech models via language-specific experts,

    T. P. Ferraz, M. Z. Boito, C. Brun, and V . Nikoulina, “Multilingual Distilwhisper: Efficient distillation of multi-task speech models via language-specific experts,” inProc. IEEE ICASSP, 2024, pp. 10 716–10 720

    2024

  25. [33]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hin- ton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” inProc. ICLR, 2017

    2017

  26. [34]

    LoRAMoE: Alleviating world knowledge forgetting in large language models via MoE-style plu- gin,

    S. Dou, E. Zhou, Y . Liu, S. Gao, W. Shen, L. Xiong, Y . Zhou, X. Wang, Z. Xi, X. Fanet al., “LoRAMoE: Alleviating world knowledge forgetting in large language models via MoE-style plu- gin,” inProc. ACL, 2024, pp. 1932–1945

    2024

  27. [35]

    Aishell-1: An open- source mandarin speech corpus and a speech recognition base- line,

    H. Bu, J. Du, X. Na, B. Wu, and H. Zheng, “Aishell-1: An open- source mandarin speech corpus and a speech recognition base- line,” inProc. O-COCOSDA, 2017, pp. 1–5

    2017

  28. [36]

    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. IEEE ICASSP, 2015, pp. 5206–5210

    2015

  29. [37]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” inProc. ICLR, 2019

    2019

  30. [38]

    Visualizing data using t-SNE

    L. Van der Maaten and G. Hinton, “Visualizing data using t-SNE.” Journal of Machine Learning Research, vol. 9, no. 11, 2008

    2008

Pith tools

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