Pith. sign in

REVIEW 5 major objections 6 minor 37 references

By assigning one expert module to each cluster of related languages, MoLGE reports lower recognition error than dense multilingual baselines on 495 languages while adding only tens of millions of trainable parameters.

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

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

T0 review · deepseek-v4-flash

2026-07-31 23:12 UTC pith:SLXSQUTN

load-bearing objection Solid empirical study of language-group MoE for multilingual ASR, but the headline claim about linguistic priors needs a better random baseline and honest parameter accounting. the 5 major comments →

arxiv 2607.24030 v1 pith:SLXSQUTN submitted 2026-07-27 cs.CL cs.SD

MoLGE: Mixture of Language Group Experts for Efficient Scaling of Massively Multilingual Speech Recognition

Sangmin Lee , Woojin Chung , Woongjib Choi , Hong-Goo Kang This is my paper
classification cs.CL cs.SD
keywords multilingual ASRmixture of expertslanguage groupinglow-rank adaptationparameter-efficient fine-tuningself-supervised speech modelscurse of multilingualitylow-resource languages
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that the 'curse of multilinguality' in speech recognition can be broken not by adding more parameters but by reorganizing them along linguistic lines. Its proposed architecture, MoLGE, assigns one shared expert module to each cluster of related languages instead of one expert per language, and applies lightweight low-rank adapters only in the upper, language-dependent layers of a frozen self-supervised speech encoder. On 495 languages it reports consistently lower character error rates than dense baselines and random grouping, with only tens of millions of additional trainable parameters. The paper further claims that explicit linguistic priors—from knowledge bases or pretrained speech embeddings—consistently beat implicit, router-learned groupings, and that the benefit is larger when the target output is orthographic scripts rather than Romanized phonetic transcripts.

Core claim

MoLGE's central claim is that grouping related languages under shared experts outperforms both dense models and per-language experts for massively multilingual ASR. Built on a frozen speech self-supervised model, MoLGE keeps the lower layers language-agnostic with shared low-rank adapters for acoustics, and places a mixture of low-rank experts in the upper layers where linguistic information lives, adding a shared expert for paralinguistic factors and an attentive statistical pooling router. The paper evaluates six grouping strategies—implicit frame/sample routing, embeddings from self-supervised and language-identification models, and geographic/genetic linguistic vectors—and reports that e

What carries the argument

The central mechanism is group-level expert routing with hierarchical low-rank adaptation. Lower layers use shared low-rank adapters for cross-lingual acoustics; upper layers use 16 language-group-specific low-rank experts chosen by an attentive statistical pooling router, plus one shared expert for speaker and paralinguistic factors. Language groups are built beforehand either implicitly (the router decides), from pretrained speech-embedding clusters, or from geographic/genetic linguistic vectors, and then assigned to experts; the paper hypothesizes that grouping acts as implicit data augmentation that stabilizes expert specialization. A two-stage dense-to-sparse upcycling initializes the m

Load-bearing premise

The load-bearing premise is that the language groups produced by linguistic or embedding-based similarity are acoustically coherent enough that one expert can serve all members; the paper itself notes that its linguistic knowledge base can place unrelated languages such as Japanese and Korean in one group, and it never ties cluster quality directly to the recognition gains.

What would settle it

Run MoLGE on the orthographic model with the same 16 clusters but with cluster labels randomly permuted across languages, holding cluster sizes and expert count fixed. If CER does not rise relative to the linguistically guided grouping, the explicit-prior claim is false and the gain is just from having multiple experts; if CER rises, the linguistic content of the groups is doing the work.

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

If this is right

  • Massively multilingual ASR can scale to hundreds of languages without billion-parameter dense models; structured language grouping is a viable alternative to uniform scaling.
  • Under a fixed number of experts, explicit linguistic priors beat implicit router-learned grouping, so pre-defined linguistic structure gives better expert specialization than learned routing alone.
  • The benefit grows with output-space complexity: orthographic-script ASR gains more from grouping than Romanized phonetic ASR, meaning language-aware design matters most where the target space is richer.
  • Grouping acts as a form of implicit data augmentation, letting low-resource languages inherit representation strength from higher-resource languages in the same group and narrowing performance gaps.
  • Going from 16 to 32 experts gives negligible gains, so a relatively small number of language-group experts can suffice at this scale.

Where Pith is reading between the lines

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

  • The paper does not directly measure cluster coherence against per-group error, so a natural next step is to test whether linguistically coherent clusters actually produce lower per-group CER; that would pin down the mechanism behind the reported gains.
  • A testable extension is to replace hard group assignments with soft or learnable group membership supervised by both language-identification signals and linguistic similarity, which might combine the strengths of implicit and explicit grouping.
  • The same group-expert architecture could plausibly transfer to speech translation or text-to-speech, where orthographic and language-specific structure matters even more than in ASR; this is not tested in the paper.
  • The 16-expert choice comes from a heuristic based on roughly 500 languages, so scaling to thousands of languages may require a data-dependent revision of the expert count rather than a fixed ratio.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes MoLGE, a method for massively multilingual ASR that adds mixture-of-LoRA-experts to a frozen speech self-supervised model, with experts assigned to language groups derived from implicit routing, SSL/LID embeddings, or URIEL+ knowledge. It evaluates on a 495-language benchmark with two 300M backbones: LAMA-UT (phonetic/Romanized output) and OmniASR (orthographic output). The main reported results are lower CER for MoLGE compared to dense and random-grouping baselines, an analysis of six grouping strategies, ablations of architectural components, and a study of expert count.

Significance. If validated, the paper offers a practical way to specialize massively multilingual ASR without greatly increasing activated parameters, and its systematic comparison of six language-grouping strategies is a useful contribution. The paper has strengths: a large-scale 495-language evaluation, two complementary ASR backbones, a code-switching analysis in Appendix C, and an honest discussion of URIEL+ imputation issues in §5.2. However, the central empirical claims currently rest on an underspecified random baseline, a possibly post-hoc selection of the best grouping method, an ambiguous parameter-efficiency statement, and a lack of repeated runs or significance testing.

major comments (5)
  1. [§4.2 / Table 1, with §3.2] The random-grouping baseline in Table 1 is not matched to the explicit groupings. §3.2 states that Ward AHC produces 'balanced and compact clusters,' but no balancing or stratification is described for the random assignment, and only one random draw is reported. If the random groups differ in language count or training hours, the 2.0-2.1 CER gaps (24.31 vs 22.27 for LAMA-UT; 25.81 vs 23.73 for OmniASR) could be caused by load imbalance or undertrained experts rather than by linguistic coherence. Provide a matched random control with identical group sizes and hours per group, run several seeds, and report mean±std.
  2. [Abstract / Table 1 vs Table 5] The abstract's 'minimal increase in trainable parameters' is contradicted by the parameter accounting. Table 1 lists ActivatedParams 343M for MoLGE vs 300M for dense, but Table 5 reports Total Params 465M for 16 experts. If the backbone is frozen in the second stage, the trainable parameters are the LoRA/router modules (165M), which is less than dense's 300M; if the backbone is also updated, the increase is 165M (55%). Either way, the paper must specify whether efficiency is measured in activated or trainable parameters, and the abstract's wording should match the actual accounting.
  3. [§5.2 / Table 2 vs Table 1] The MoLGE rows in Table 1 appear to correspond to the best-performing grouping strategy in Table 2 (LID for LAMA-UT, Geographic for OmniASR), chosen on the same test benchmark. No held-out validation or model-selection protocol is described. This raises the risk that the reported gains over dense and random baselines rely on oracle selection among six strategies. Report all six strategies against the same baselines, or select the grouping method on a validation set and describe the procedure explicitly.
  4. [§2.2-2.3 / §5.1] No empirical comparison to prior language-expert/MoLE methods is provided, despite the related work discussing Kwon & Chung 2023, Bagat et al. 2025, Li et al. 2025b, and Mu et al. 2025. Without these baselines, the paper cannot support its positioning that group-level experts improve over language-specific experts, especially because Table 5 shows that increasing experts from 16 to 32 gives no gain. Add at least one representative language-expert/LoRA baseline for each backbone.
  5. [Tables 1-5] There are no error bars, repeated seeds, or significance tests anywhere in the experimental section. The word 'consistently' in the abstract is not supported by the evidence presented; a single run per configuration cannot establish consistency. At least for the central dense-vs-MoLGE comparison, report multiple seeds (or a significance test) and per-language confidence intervals, not just aggregate CER.
minor comments (6)
  1. [Figure 3] Specify how the confidence intervals are computed and provide confidence intervals for the correlation coefficients r (−0.24 vs −0.26).
  2. [Table 3 caption] Define 'M' (pure MoLE) and 'L' (vanilla LoRA) in the caption rather than only in the text.
  3. [§3.2] The k ≈ √(N/2) rule is attributed to Milligan & Cooper (1985), but that reference does not state this rule in the cited form. Provide a specific page or describe it as an ad-hoc heuristic.
  4. [Table 5] Clarify that the CER 24.51 is for sample-level implicit grouping, and state whether the total-parameter counts include the frozen backbone.
  5. [Table 4] Add group identifiers to show that the listed languages are actually in the same group, since the text claims this but the table does not show group membership.
  6. [§3.1] Minor typographical inconsistency: 'MOLE' should be 'MoLE'.

Circularity Check

0 steps flagged

No significant circularity; MoLGE's reported gains are empirical and not forced by construction.

full rationale

The paper's central claims rest on measured CER differences in Table 1 and Table 2, comparing dense, random-group, and explicitly grouped MoLGE variants on 495 languages. The grouping strategies are built from external resources (URIEL+ vectors, OmniASR-W2V embeddings, MMS-LID embeddings) rather than from the evaluation metric, so the improvement is an empirical result rather than an identity. No equation in the paper reduces MoLGE's CER to the grouping labels or to a fitted constant; the routing loss is weighted at 0.01 relative to CTC, so ASR performance is not directly optimized to match the grouping prior. The one mildly self-referential element is the reported routing accuracy for explicit grouping: Section 4.3 states that 'explicit grouping uses cross-entropy for supervised group prediction,' so high routing accuracy partly reflects the router's own training objective. However, the paper's main conclusion is supported by CER, not by routing accuracy, and the comparison between random and linguistically coherent labels still has empirical content. The use of LAMA-UT, a prior system by overlapping authors, as a backbone is a minor self-citation, but it is not load-bearing because the same conclusions hold on the independently sourced OmniASR backbone. Concerns about the underspecified random baseline and selection of the best grouping strategy on the same benchmark are validity and generalization risks, not circular reasoning. The paper also openly acknowledges limitations such as URIEL+ imputation grouping isolates (Section 5.2) and the impracticality of full phonetic transcription (Appendix B), which is consistent with an empirical rather than circular derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

These are the main hand-set choices and prior assumptions the reported performance depends on. None are derived from first principles; the most consequential is the layer-wise specialization assumption (axiom 1) and the validity of URIEL+ imputed distances for low-resource languages (axiom 3).

free parameters (5)
  • number_of_experts = 16
    Set by sqrt(500/2) heuristic and pilot study comparing 8/16/32 on the same benchmark (Sec 3.2, Table 5).
  • LoRA_rank = 32
    Fixed for all experts and shared modules; no sensitivity analysis reported (Sec 4.3).
  • routing_loss_weight = 0.01
    Weight of routing loss relative to CTC loss (Sec 4.3); chosen by hand without ablation.
  • sampling_exponents_beta = beta_L = beta_D = 0.5
    Multilevel sampling exponents from Pratap et al./Keren et al.; affects language balance (Sec 4.3).
  • segmentation_thresholds = 30s; 150 chars; per-language alignment scores
    Omnilingual corpus preprocessing: 30-second chunks, forced segmentation at ~150 chars, per-language alignment-score pruning (Appendix A); thresholds not ablated.
axioms (7)
  • domain assumption Lower S3M layers encode language-agnostic acoustic features, higher layers encode language-specific linguistic information (Pasad et al. 2021, 2023).
    Used to justify shared LoRA in lower layers and MoLE only in upper layers (Sec 3.1). If false, the hierarchical allocation loses its basis.
  • domain assumption Related languages share phonological, morphological, and genealogical structure that improves shared expert training.
    Central hypothesis of the paper (Sec 1, 3.2); the grouping benefit depends on this.
  • domain assumption URIEL+ vectors and their imputation for low-resource/isolated languages produce meaningful distance matrices.
    Knowledge-guided groups (geographic/genetic) are built from these vectors (Sec 3.2); authors note isolates/dialects can be misassigned (Sec 5.2).
  • standard math Ward's hierarchical clustering yields coherent, balanced language clusters.
    Clustering algorithm choice for embedding- and knowledge-guided grouping; no alternative clustering compared (Sec 3.2).
  • ad hoc to paper The k≈sqrt(N/2) heuristic gives a near-optimal number of experts.
    Used to fix 16 experts; later confirmed by a pilot study on the same benchmark (Sec 3.2, Table 5).
  • domain assumption Two-stage dense-to-sparse upcycling preserves dense representations while adding expert specialization.
    Training strategy adapted from LLM upcycling (Sec 3.3); no comparison to training from scratch.
  • domain assumption Romanized transcription is a valid phonetic proxy for evaluating phonetic modeling at 495-language scale.
    Justifies the LAMA-UT phonetic-domain experiments (Appendix B); explicitly acknowledged as not true IPA.

pith-pipeline@v1.3.0-alltime-deepseek · 15711 in / 13686 out tokens · 111031 ms · 2026-07-31T23:12:50.942792+00:00 · methodology

0 comments
read the original abstract

Massively multilingual automatic speech recognition (ASR) models covering hundreds of languages must maintain robust performance across diverse linguistic and acoustic conditions. However, these models often encounter the curse of multilinguality, where model capacity is diluted across languages. To address this challenge, we propose Mixture of Language Group Experts (MoLGE), built upon speech self-supervised models (S3Ms). MoLGE assigns dedicated expert modules to clusters of similar languages, reducing the number of required submodules compared to conventional language-specific Mixture-of-Experts (MoE) schemes. It further integrates a hierarchical Low-Rank Adaptation (LoRA) strategy into the disentangled acoustic and linguistic components of the S3M architecture, enabling efficient modeling of language-specific characteristics while maintaining parameter efficiency. Further, we investigate the impact of language grouping strategies based on both linguistic and data-driven criteria on overall performance, providing an interpretable perspective on how language structure influences scalability in multilingual speech systems. In experiments, we evaluate MoLGE on a multilingual benchmark encompassing 495 languages. Results demonstrate that MoLGE consistently outperforms dense multilingual baselines with a minimal increase in trainable parameters. Notably, these language grouping strategies yield substantial improvements for both phonetic and orthographic aspects of ASR modeling. Our findings suggest that structured language specialization provides an effective pathway for massively scaling language coverage of multilingual ASR.

Figures

Figures reproduced from arXiv: 2607.24030 by Hong-Goo Kang, Sangmin Lee, WooJin Chung, Woongjib Choi.

Figure 1
Figure 1. Figure 1: Overall architecture of MoLGE. The conventional S3M is decomposed into lower [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the overall procedure for language grouping, covering implicit, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Regression plot between CER and training data. Each point indicates the mean [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of language clustering across different grouping strategies. Each [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Duration distributions in the omnilingual ASR corpus, aggregated across all splits. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of embedding-guided clusters. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of knowledge-guided clusters. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 2 canonical work pages

  1. [2]

    XLS-R: Self-supervised cross-lingual speech representation learning at scale

    Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. XLS-R: Self-supervised cross-lingual speech representation learning at scale. InInterspeech 2022, pp. 2278–2282,

    2022

  2. [3]

    Mixture of LoRA experts for low- resourced multi-accent automatic speech recognition

    Rapha¨el Bagat, Irina Illina, and Emmanuel Vincent. Mixture of LoRA experts for low- resourced multi-accent automatic speech recognition. InInterspeech 2025, pp. 1143–1147,

    2025

  3. [5]

    Unsupervised cross-lingual representation learning for speech recognition

    Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, and Michael Auli. Unsupervised cross-lingual representation learning for speech recognition. In Interspeech 2021, pp. 2426–2430,

    2021

  4. [6]

    FLEURS: Few-shot learning evaluation of universal representations of speech

    11 Published as a conference paper at COLM 2026 Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. FLEURS: Few-shot learning evaluation of universal representations of speech. InIEEE Spoken Language Technology Workshop, pp. 798–805. IEEE,

    2026

  5. [7]

    Each plot is annotated with ISO 639-3 language codes following the Ethnologue4 convention

    Languages assigned to the same group are indicated by the same color. Each plot is annotated with ISO 639-3 language codes following the Ethnologue4 convention. 4https://www.ethnologue.com/ 18 Published as a conference paper at COLM 2026 (a) Cluster formation using the SSL embedding. (b) Cluster formation using the LID embedding. Figure 6: Comparison of e...

    2026

  6. [11]

    Mahadeva Prasanna, Deepu Vijayasenan, and Sriram Ganapathy

    Shareef Babu Kalluri, Prachi Singh, Pratik Roy Chowdhuri, Apoorva Kulkarni, Shikha Baghel, Pradyoth Hegde, Swapnil Sontakke, Deepak K T, S.R. Mahadeva Prasanna, Deepu Vijayasenan, and Sriram Ganapathy. The Second DISPLACE Challenge: DIarization of SPeaker and LAnguage in Conversational Environments. InInterspeech 2024, pp. 1630– 1634,

    2024

  7. [12]

    Siva Kalyan and Alexandre Franc ¸ois

    doi: 10.21437/Interspeech.2024-1833. Siva Kalyan and Alexandre Franc ¸ois. Freeing the comparative method from the tree model: A framework for historical glottometry.Senri Ethnological Studies, 98:59–89,

    work page doi:10.21437/interspeech.2024-1833 2024

  8. [14]

    Group then scale: Dy- namic mixture-of-experts multilingual language model

    13 Published as a conference paper at COLM 2026 Chong Li, Yingzhuo Deng, Jiajun Zhang, and Chengqing Zong. Group then scale: Dy- namic mixture-of-experts multilingual language model. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 1730–1754, 2025a. Jiahong Li, Yiwen Shao, Jianheng Zhuo, Chenda Li, Liliang Tang, Dong Yu, and Yanm...

    2026

  9. [16]

    A first south african corpus of multilingual code-switched soap opera speech

    Thomas Niesler et al. A first south african corpus of multilingual code-switched soap opera speech. InProceedings of the eleventh international conference on language resources and evaluation (LREC 2018),

    2018

  10. [17]

    Attentive statistics pooling for deep speaker embedding

    Koji Okabe, Takafumi Koshinaka, and Koichi Shinoda. Attentive statistics pooling for deep speaker embedding. InInterspeech 2018, pp. 2252–2256,

    2018

  11. [19]

    OWSM v4: Improving open Whisper-style speech models via data scaling and cleaning

    Yifan Peng, Muhammad Shakeel, Yui Sudo, William Chen, Jinchuan Tian, Chyi-Jiunn Lin, and Shinji Watanabe. OWSM v4: Improving open Whisper-style speech models via data scaling and cleaning. InInterspeech 2025, pp. 2225–2229,

    2025

  12. [20]

    RomanLens: The role of latent Romanization in multilinguality in LLMs

    Alan Saji, Jaavid Aktar Husain, Thanmay Jayakumar, Raj Dabre, Anoop Kunchukuttan, and Ratish Puduppully. RomanLens: The role of latent Romanization in multilinguality in LLMs. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 26410– 26429,

    2025

  13. [21]

    ML-SUPERB: Multilingual Speech Universal PERformance Benchmark

    Jiatong Shi, Dan Berrebbi, William Chen, En-Pei Hu, Wei-Ping Huang, Ho-Lam Chung, Xu- ankai Chang, Shang-Wen Li, Abdelrahman Mohamed, Hung yi Lee, and Shinji Watanabe. ML-SUPERB: Multilingual Speech Universal PERformance Benchmark. InInterspeech 2023, pp. 884–888,

    2023

  14. [22]

    Jiatong Shi, Shih-Heng Wang, William Chen, Martijn Bartelds, Vanya Bannihatti Kumar, Jinchuan Tian, Xuankai Chang, Dan Jurafsky, Karen Livescu, Hung yi Lee, and Shinji Watanabe

    doi: 10.21437/Interspeech.2023-1316. Jiatong Shi, Shih-Heng Wang, William Chen, Martijn Bartelds, Vanya Bannihatti Kumar, Jinchuan Tian, Xuankai Chang, Dan Jurafsky, Karen Livescu, Hung yi Lee, and Shinji Watanabe. ML-SUPERB 2.0: Benchmarking Multilingual Speech Models Across Modeling Constraints, Languages, and Datasets. InInterspeech 2024, pp. 1230–1234,

    doi:10.21437/interspeech.2023-1316 2023

  15. [23]

    Igor Sz¨oke, Miroslav Sk´acel, Ladislav Moˇsner, Jakub Paliesek, and Jan ˇCernock`y

    doi: 10.21437/Interspeech.2024-2248. Igor Sz¨oke, Miroslav Sk´acel, Ladislav Moˇsner, Jakub Paliesek, and Jan ˇCernock`y. Building and evaluation of a real room impulse response dataset.IEEE Journal of Selected Topics in Signal Processing, 13(4):863–876,

    work page doi:10.21437/interspeech.2024-2248 2024

  16. [25]

    k2SSL: A faster and better framework for self-supervised speech representation learning

    15 Published as a conference paper at COLM 2026 Yifan Yang, Jianheng Zhuo, Zengrui Jin, Ziyang Ma, Xiaoyu Yang, Zengwei Yao, Liyong Guo, Wei Kang, Fangjun Kuang, Long Lin, et al. k2SSL: A faster and better framework for self-supervised speech representation learning. InIEEE International Conference on Multimedia and Expo, pp. 1–6. IEEE,

    2026

  17. [26]

    mHuBERT-147: A compact multilingual HuBERT model

    Marcely Zanon Boito, Vivek Iyer, Nikolaos Lagos, Laurent Besacier, and Ioan Calapodescu. mHuBERT-147: A compact multilingual HuBERT model. InInterspeech 2024, pp. 3939– 3943,

    2024

  18. [27]

    Linear- complexity self-supervised learning for speech processing

    Shucong Zhang, Titouan Parcollet, Rogier van Dalen, and Sourav Bhattacharya. Linear- complexity self-supervised learning for speech processing. InInterspeech 2024, pp. 3480– 3484,

    2024

  19. [28]

    Google USM: Scaling automatic speech recognition beyond 100 languages.arXiv preprint arXiv:2303.01037,

    Yu Zhang, Wei Han, James Qin, Yongqiang Wang, Ankur Bapna, Zhehuai Chen, Nanxin Chen, Bo Li, Vera Axelrod, Gary Wang, et al. Google USM: Scaling automatic speech recognition beyond 100 languages.arXiv preprint arXiv:2303.01037,

    Pith/arXiv arXiv

  20. [29]

    Efficiently democratizing medical LLMs for 50 languages via a mixture of language family experts.arXiv preprint arXiv:2410.10626,

    Guorui Zheng, Xidong Wang, Juhao Liang, Nuo Chen, Yuping Zheng, and Benyou Wang. Efficiently democratizing medical LLMs for 50 languages via a mixture of language family experts.arXiv preprint arXiv:2410.10626,

    Pith/arXiv arXiv

  21. [30]

    Opportunities and challenges of large language models for low-resource languages in humanities research.arXiv preprint arXiv:2412.04497,

    Tianyang Zhong, Zhenyuan Yang, Zhengliang Liu, Ruidong Zhang, Weihang You, Yiheng Liu, Haiyang Sun, Yi Pan, Yiwei Li, Yifan Zhou, et al. Opportunities and challenges of large language models for low-resource languages in humanities research.arXiv preprint arXiv:2412.04497,

    Pith/arXiv arXiv

  22. [31]

    URL https://arxiv

    doi: 10.48550/ARXIV .2204.03067. URL https://arxiv. org/abs/2204.03067. Jian Zhu, Farhan Samir, Eleanor Chodroff, and David R Mortensen. Zipa: A family of efficient models for multilingual phone recognition. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 19568– 19585,

    Pith/arXiv arXiv doi:10.48550/arxiv.2204.03067

  23. [32]

    (b) Duration of the dataset after segmentation

    16 Published as a conference paper at COLM 2026 (a) Duration of the original dataset. (b) Duration of the dataset after segmentation. Figure 5: Duration distributions in the omnilingual ASR corpus, aggregated across all splits. A Dataset Details and Preprocessing Strategy A.1 Dataset Details FLEURS.FLEURS is a multilingual dataset encompassing 102 languag...

    2026

  24. [33]

    Rather than discarding them, which would significantly reduce the data, we apply a segmentation strategy to preserve the data capacity. Segmentation Strategy.To address the aforementioned issue, we adopt a three-stage segmentation pipeline for utterances longer than 30 seconds: (1) segmentation using hard punctuation (e.g., periods, question marks, exclam...

    2023

  25. [34]

    However, constructing reliable phonetic transcriptions at the scale considered in this work (495 languages) remains impractical

    and PRiSM (Bharadwaj et al., 2026). However, constructing reliable phonetic transcriptions at the scale considered in this work (495 languages) remains impractical. The primary challenge lies in multilingual grapheme-to-phoneme (G2P) conversion. Exist- ing multilingual G2P toolkits provide only limited language coverage. For example, Phone- mizer (Bernard...

    2026

  26. [35]

    Consequently, generating consistent, high-quality phoneme transcriptions across hundreds of languages is currently infeasible

    rely extensively on zero-shot G2P generation, making phonetic quality diffi- cult to guarantee for many low-resource languages. Consequently, generating consistent, high-quality phoneme transcriptions across hundreds of languages is currently infeasible. Romanization provides a practical alternative that can be consistently applied across a substantially ...

    2024

  27. [36]

    In contrast, frame-level routing does not exhibit a clear or consistent language-dependent grouping pattern

    The analysis indicates that, under sentence-level routing, the assigned expert group is largely de- termined by the matrix language, with agreement rates of 78.29% and 69.91% on DISPLACE 2024 and the South African Soap Opera corpus, respectively. In contrast, frame-level routing does not exhibit a clear or consistent language-dependent grouping pattern. T...

    2024

  28. [2012]

    Multilingual large language models and curse of multilinguality.arXiv preprint arXiv:2406.10602,

    Daniil Gurgurov, Tanja B¨aumel, and Tatiana Anikina. Multilingual large language models and curse of multilinguality.arXiv preprint arXiv:2406.10602,

    Pith/arXiv arXiv

  29. [2013]

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

    Wenxuan Wang, Guodong Ma, Yuke Li, and Binbin Du. Language-routing mixture of experts for multilingual and code-switching speech recognition. InInterspeech 2023, pp. 1389–1393,

    2023

  30. [2018]

    Omnilingual ASR: Open-source multilingual speech recognition for 1600+ languages.arXiv preprint arXiv:2511.09690,

    Gil Keren, Artyom Kozhevnikov, Yen Meng, Christophe Ropers, Matthew Setzler, Skyler Wang, Ife Adebara, Michael Auli, Can Balioglu, et al. Omnilingual ASR: Open-source multilingual speech recognition for 1600+ languages.arXiv preprint arXiv:2511.09690,

    arXiv

  31. [2019]

    Sustainable self-supervised learning for speech represen- tations.arXiv preprint arXiv:2406.07696,

    Luis Lugo and Valentin Vielzeuf. Sustainable self-supervised learning for speech represen- tations.arXiv preprint arXiv:2406.07696,

    Pith/arXiv arXiv

  32. [2020]

    Alabi, Xuechen Liu, Dietrich Klakow, and Junichi Yamagishi

    Jesujoba O. Alabi, Xuechen Liu, Dietrich Klakow, and Junichi Yamagishi. AfriHuBERT: A self-supervised speech representation model for African languages. InInterspeech 2025, pp. 4023–4027,

    2025

  33. [2021]

    Prism: Benchmarking phone realization in speech models.arXiv preprint arXiv:2601.14046,

    Shikhar Bharadwaj, Chin-Jou Li, Yoonjae Kim, Kwanghee Choi, Eunjung Yeo, Ryan Soh-Eun Shim, Hanyu Zhou, Brendon Boldt, Karen Rosero Jacome, Kalvin Chang, et al. Prism: Benchmarking phone realization in speech models.arXiv preprint arXiv:2601.14046,

    Pith/arXiv arXiv

  34. [2022]

    Dynamic language group-based MoE: Enhancing code-switching speech recognition with hierarchical routing

    12 Published as a conference paper at COLM 2026 Hukai Huang, Shenghui Lu, Yahui Shan, He Qu, Fengrun Zhang, Wenhao Guan, Qingyang Hong, and Lin Li. Dynamic language group-based MoE: Enhancing code-switching speech recognition with hierarchical routing. InProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 1–5. IEEE,

    2026

  35. [2023]

    14 Published as a conference paper at COLM 2026 Juan N Pava, Caroline Meinhardt, Haifa Badi Uz Zaman, Toni Friedman, Sang T Truong, Daniel Zhang, Vukosi Marivate, and Sanmi Koyejo. Mind the (language) gap: Mapping the challenges of LLM development in low-resource language contexts.Stanford Institute for Human-Centered Artificial Intelligence (HAI): Stanfo...

    2026

  36. [2024]

    Upcycling large language models into mixture of experts.arXiv preprint arXiv:2410.07524,

    Ethan He, Abhinav Khattar, Ryan Prenger, Vijay Korthikanti, Zijie Yan, Tong Liu, Shiqing Fan, Ashwath Aithal, Mohammad Shoeybi, and Bryan Catanzaro. Upcycling large language models into mixture of experts.arXiv preprint arXiv:2410.07524,

    Pith/arXiv arXiv

  37. [2025]

    Mixtral of experts.arXiv preprint arXiv:2401.04088,

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts.arXiv preprint arXiv:2401.04088,

    Pith/arXiv arXiv