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

# MoLGE: Mixture of Language Group Experts for Efficient Scaling of Massively Multilingual Speech Recognition
Source: [https://arxiv.org/html/2607.24030](https://arxiv.org/html/2607.24030)
Sangmin Lee, Woojin Chung, Woongjib Choi, and Hong\-Goo Kang Dept\. of Electronics and Electrical Engineering, Yonsei University, Seoul, South Korea \{sangmin\_lee,woojinchung,woongzip1\}@dsp\.yonsei\.ac\.kr, hgkang@yonsei\.ac\.kr

###### 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 thecurse 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\.

## 1Introduction

Scaling automatic speech recognition \(ASR\) to hundreds of languages within a unified model remains a core challenge in multilingual speech processing\. While these systems aim for broad accessibility, they must navigate significant linguistic and acoustic diversity\. Consequently, performance often degrades as the number of supported languages increases, a phenomenon primarily driven by limited model capacity relative to the task’s complexity\.

Since the advent of speech self\-supervised models \(S3Ms\)\(Baevskiet al\.,[2020](https://arxiv.org/html/2607.24030#bib.bib1); Hsuet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib2); Chenet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib3)\), speech representation learning has advanced significantly by leveraging large\-scale unlabeled corpora\. These models have achieved robust performance across various downstream tasks, including ASR, in some cases surpassing human\-level transcription accuracy on monolingual benchmarks\. Consequently, recent research has increasingly focused on extending S3Ms to multilingual settings\(Conneauet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib4); Babuet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib5); Zanon Boitoet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib6); Pratapet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib7); Chenet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib8)\)\. Current approaches primarily rely on scaling training data to millions of hours\(Zhanget al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib11)\)and increasing model capacity to billions of parameters\(Kerenet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib9)\)to improve multilingual coverage and cross\-lingual generalization\.

Despite these advances, significant challenges remain\. Increasing dataset size and model capacity is computationally demanding and often impractical at scale\(Zhanget al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib20); Yanget al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib21)\)\. Recent massively multilingual systems already require hundreds of millions to billions of parameters, rendering both pretraining and adaptation resource\-intensive, particularly in computationally constrained settings\(Lugo and Vielzeuf,[2024](https://arxiv.org/html/2607.24030#bib.bib22)\)\.

Furthermore, multilingual speech models are typically trained on highly imbalanced datasets dominated by high\-resource languages\. This imbalance exacerbates the curse of multilinguality\(Conneauet al\.,[2020](https://arxiv.org/html/2607.24030#bib.bib16); Gurgurovet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib17)\), as fixed model capacity is shared across many languages while being disproportionately optimized for a small subset\. Consequently, a single dense multilingual model often underperforms relative to specialized monolingual systems\(Wanget al\.,[2020](https://arxiv.org/html/2607.24030#bib.bib18); Chenet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib19)\), highlighting the limitations of uniform scaling for multilingual speech modeling\. These limitations reveal a fundamental tension between language scalability and resource accessibility\. While scaling multilingual models aims to support underrepresented languages, their increasing computational demands driven by model size and training data limit their practical utility in the settings they are intended to serve\(Zhonget al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib23); Cahyawijaya,[2024](https://arxiv.org/html/2607.24030#bib.bib24); Pavaet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib25)\)\.

To this end, we argue that this tension can be alleviated by shifting from uniform capacity scaling to structured architectural design informed by language relationships\. Languages are not independent entities; they share phonological, morphological, and genealogical properties that define a structured space of linguistic similarity\(Bloomfield,[1983](https://arxiv.org/html/2607.24030#bib.bib50); Littellet al\.,[2017](https://arxiv.org/html/2607.24030#bib.bib54)\)\. Exploiting this structure as an architectural prior allows model capacity to be allocated in a linguistically principled manner, providing a more efficient pathway for multilingual scaling than exhaustive parameter or data expansion\.

Building on this motivation, we proposeMixture of Language Group Experts \(MoLGE\), a parameter\-efficient and language\-aware framework for massively multilingual speech modeling\. MoLGE enables structured language specialization by incorporating linguistic relationships directly into its expert allocation mechanism, without relying solely on scaling\.

Specifically, our framework is built on two design principles\. First, motivated by evidence that lower S3M layers capture universal acoustic features while higher layers encode language\-specific information\(Pasadet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib12);[2023](https://arxiv.org/html/2607.24030#bib.bib13)\), we adopt a speech\-oriented adaptation strategy\. This includes hierarchical adaptation, where shared LoRA modules model universal acoustics in lower layers and language group\-specific experts capture linguistic characteristics in higher layers, as well as shared experts and an attentive pooling router to improve acoustic modeling and stabilize expert specialization\.

Second, we explore language grouping as a mechanism for injecting linguistic prior knowledge into the model architecture\. We hypothesize that linguistically or acoustically similar languages share structural patterns, and that grouping them under shared experts effectively increases the accessible training signal for each expert, acting as a form of implicit data augmentation\. To this end, we investigate six grouping strategies spanning implicit, embedding\-guided, and knowledge\-guided approaches, providing a comprehensive evaluation of how different types of linguistic prior influence expert specialization\.

Beyond the architectural design, we conduct a systematic analysis of when and why language\-aware modeling is most beneficial\. By evaluating MoLGE across two ASR models with distinct output characteristics, phonetic \(language\-agnostic Romanized transcription\) and orthographic \(language\-specific scripts\), we examine how the complexity of the target linguistic space modulates the effectiveness of structured language grouping\.

Our contributions are summarized as follows:

- •We propose MoLGE, an efficient and scalable framework for massively multilingual ASR that incorporates language structure as an architectural prior for expert specialization\.
- •We integrate speech modeling principles into LoRA and MoLE, incorporating hierarchical adaptation, shared experts, and an attentive pooling router to improve acoustic modeling and expert specialization\.
- •We provide a systematic study of six language grouping strategies across 495 languages, showing that explicit linguistic priors consistently outperform implicit or random priors, and that their benefit scales with the complexity of the target linguistic space\.

## 2Related Work

### 2\.1Multilingual Speech Foundation Models

Multilingual speech foundation models have primarily advanced by scaling model capacity and training data to support an increasing number of languages\. Early research, such as XLSR\-53\(Conneauet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib4)\), demonstrated that multilingual self\-supervised pretraining with wav2vec 2\.0\(Baevskiet al\.,[2020](https://arxiv.org/html/2607.24030#bib.bib1)\)improves cross\-lingual transfer and generalization for ASR\. Subsequent models, including XLS\-R\(Babuet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib5)\), MMS\(Pratapet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib7)\), mHuBERT\-147\(Zanon Boitoet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib6)\), AfriHuBERT\(Alabiet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib28)\), XEUS\(Chenet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib8)\), and Omnilingual ASR\(Kerenet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib9)\), have extended this paradigm by scaling to hundreds or thousands of languages and increasing model capacity to the billion\-parameter regime\. In parallel, large\-scale supervised Whisper\-style approaches\(Radfordet al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib10); Penget al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib29)\)have achieved robust multilingual ASR performance by training on massive labeled datasets\. Despite these advances, existing methods predominantly rely on uniform scaling of model size and training data, with limited attention to language\-aware modeling that explicitly accounts for linguistic diversity\.

### 2\.2Scalable Parameter\-Efficient Finetuning

Parameter\-efficient fine\-tuning \(PEFT\) methods have been widely adopted to adapt large pretrained models without updating the entire parameter set\. Approaches such as adapters\(Houlsbyet al\.,[2019](https://arxiv.org/html/2607.24030#bib.bib26); Pfeifferet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib27)\)and Low\-Rank Adaptation \(LoRA\)\(Huet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib14)\)enable efficient task\-specific adaptation by introducing a minimal number of trainable parameters\. These methods have been successfully applied across both natural language processing \(NLP\) and speech domains to maintain performance while reducing computational overhead\. In multilingual speech modeling, PEFT has been used to scale models across numerous languages\. For example,Pratapet al\.\([2024](https://arxiv.org/html/2607.24030#bib.bib7)\)employs language\-specific adapters to extend dense multilingual models to thousands of languages\. However, assigning individual adapters to every language becomes impractical as the scale increases\.

To improve scalability, Mixture\-of\-Experts \(MoE\)\(Jacobset al\.,[1991](https://arxiv.org/html/2607.24030#bib.bib30)\)approaches introduce conditional computation by activating only a subset of parameters for each input\. Recent research further integrates MoE with PEFT, such as Mixture of LoRA Experts \(MoLE\)\(Wuet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib15)\), enabling expert specialization through lightweight adaptation modules\. Nevertheless, existing MoE\-based multilingual methods typically allocate experts at the granularity of individual languages\(Gauret al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib32); Kwon and Chung,[2023](https://arxiv.org/html/2607.24030#bib.bib33); Muet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib34)\), which limits their scalability in massively multilingual scenarios\.

Recent studies in NLP have begun exploring language grouping within MoE frameworks\(Zhenget al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib35); Liet al\.,[2025a](https://arxiv.org/html/2607.24030#bib.bib36); Lianget al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib37); Janeiroet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib38)\), suggesting that structured parameter sharing across related languages improves scalability\. However, such approaches remain underexplored in massively multilingual speech modeling\. Existing methods typically apply language\-specific LoRA modules to only a small subset of languages\(Wanget al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib39); Bagatet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib40); Liet al\.,[2025b](https://arxiv.org/html/2607.24030#bib.bib41)\)or focus on specialized scenarios such as code\-switching\(Maet al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib42); Huanget al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib43)\), limiting their applicability to large and diverse language sets\.

### 2\.3Position of Our Work

Our work departs from prior approaches in two key dimensions\.

First, instead of relying on uniform scaling, we integrate PEFT with speech\-aware adaptation and language\-group\-based routing to enable structured, scalable specialization for a massive set of languages\. Second, while prior group\-aware routing has primarily focused on textual inputs, we address the unique challenges of language grouping from a speech\-oriented perspective by conducting an in\-depth analysis of various language grouping strategies\.

Taken together, we present an efficient scaling method for massively multilingual speech recognition, along with the first systematic study of language grouping and its impact\.

![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/molge_overall_arch4.png)Figure 1:Overall architecture of MoLGE\. The conventional S3M is decomposed into lower acoustic and upper linguistic layers, each with distinct adaptation strategies\.

## 3Proposed Method

We investigate the role of language grouping in multilingual speech recognition through two primary hypotheses\. First, grouping linguistically or acoustically similar languages enables experts to capture common structural characteristics, improving parameter efficiency and representation quality, particularly in massively multilingual settings\. Second, under a fixed model capacity \(i\.e\., a constrained number of experts\), explicit language grouping strategies based on linguistic or statistical priors facilitate more effective expert specialization than implicit, router\-driven grouping learned solely from data\.

### 3\.1Mixture of Language Group Experts

In this section, we presentMixture of Language Group Experts \(MoLGE\), a speech\-centric MoLE framework designed for efficient expert specialization through structured language grouping\. In contrast to conventional approaches that allocate experts at the individual language level, MoLGE clusters related languages and assigns shared experts to each group\. This strategy significantly improves scalability in massively multilingual settings\. The overall architecture of MoLGE is illustrated in Figure[1](https://arxiv.org/html/2607.24030#S2.F1)\.

We adopt a hierarchical modeling approach based on the layer\-wise functional specialization of speech encoders, where lower layers primarily encode language\-agnostic acoustic features and higher layers capture language\-dependent linguistic representations\(Pasadet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib12);[2023](https://arxiv.org/html/2607.24030#bib.bib13)\)\. Following this observation, we applyshared LoRAmodules to the lower layers to model global acoustic variability\. Since acoustic characteristics across human speech are largely transferable, a unified parameter space efficiently captures cross\-lingual acoustic factors without introducing routing overhead\. Conversely, we deployMixture of LoRA Experts \(MoLE\)exclusively in the upper layers\. This design allocates expert capacity to the linguistic representation space, where language\-specific disambiguation is most critical, enabling efficient modeling of diverse multilingual environments\.

To further adapt the MoLE framework for speech recognition, we introduce two key modifications\. Unlike text, speech signals encode both linguistic content and paralinguistic information, such as speaking rate and speaker identity\. We introduce ashared expertto capture these paralinguistic factors, which remain largely invariant across different languages\. This allows the language\-specific experts to focus more effectively on distinct linguistic representations\. In addition, we employ anAttentive Statistical Pooling \(ASP\) classifier\(Okabeet al\.,[2018](https://arxiv.org/html/2607.24030#bib.bib44)\)as the routing mechanism\. Given that speech signals often contain temporally imbalanced information, such as long silences, ASP utilizes attention\-weighted temporal pooling to emphasize salient acoustic regions while suppressing irrelevant frames\. This ensures the router identifies more informative representations for expert assignment\.

![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/molge_grouping2.png)Figure 2:Illustration of the overall procedure for language grouping, covering implicit, embedding\-guided, and knowledge\-guided strategies\.
### 3\.2Language Grouping Strategy

In this section, we construct six language grouping strategies based on diverse information sources\. We consider three categories:implicit,embedding\-guided, andknowledge\-guided, each comprising two variants\. The detailed grouping procedure is illustrated in Figure[2](https://arxiv.org/html/2607.24030#S3.F2)\.

Following the principles of statistical cluster analysis\(Milligan and Cooper,[1985](https://arxiv.org/html/2607.24030#bib.bib45)\), the optimal number of clusters can be approximated ask≈N/2k\\approx\\sqrt\{N/2\}, whereNNdenotes the number of data points\. Given our experimental setting of approximately 500 languages \(N≈500N\\approx 500\), this heuristic suggestsk≈15\.8k\\approx 15\.8\. Accordingly, we fix the number of experts to 16 across all strategies\. This choice provides a practical trade\-off between model capacity and computational efficiency in massively multilingual scenarios\.

Implicit Grouping\.Implicit grouping refers to a strategy in which languages are not assigned to experts via predefined labels; instead, groupings emerge naturally from the learned behavior of the router\. This represents the foundational form of MoE and has been widely studied in the literature\(Lepikhinet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib46); Feduset al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib47); Jianget al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib31)\)\.

Under this strategy, we evaluate two distinct routing granularities: frame\-level and utterance\-level routing\(Fanet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib48)\)\. Frame\-level routing treats each timestep of the speech embedding as an individual token, operating analogously to token\-level routing in MoE models in NLP\. It allows for fine\-grained, dynamic expert assignment throughout the duration of a signal\. In contrast, utterance\-level routing treats an entire utterance as a single unit, routing the complete sequence to a specific expert via a shared router\. This ensures consistency in expert application across the temporal dimension of the speech input\.

Embedding\-Guided Grouping\.Embedding\-guided grouping derives language clusters from high\-dimensional representations generated by pretrained speech foundation models\. Under this strategy, we first extract speech embeddings and construct language prototype vectors by aggregating these representations at the language level\. Then, we apply Agglomerative Hierarchical Clustering \(AHC\) with Ward’s linkage\(Ward Jr,[1963](https://arxiv.org/html/2607.24030#bib.bib49)\)to partition the languages\. This clustering approach is motivated by the inherently hierarchical and phylogenetic structure of human languages\(Bloomfield,[1983](https://arxiv.org/html/2607.24030#bib.bib50); François,[2015](https://arxiv.org/html/2607.24030#bib.bib51); Heggartyet al\.,[2010](https://arxiv.org/html/2607.24030#bib.bib52); Kalyan and François,[2018](https://arxiv.org/html/2607.24030#bib.bib53)\)\. Moreover, Ward’s method minimizes within\-cluster variance to encourage the formation of balanced and compact clusters\. This structural property is critical for ensuring stable load balancing and mitigating collapse during training\.

To evaluate the impact of different representational biases, we implement this strategy using two distinct pretrained speech models\. The first variant utilizes OmniASR\-W2V111https://huggingface\.co/facebook/omniASR\-W2V\-1B, a multilingual wav2vec2\.0\-style model trained on 1,239 languages\. Because this model is optimized with a general self\-supervised learning \(SSL\) objective, it captures diverse speech characteristics, including speaker, phonetic, and language\-level information\. The second variant employs MMS\-LID\-4017222https://huggingface\.co/facebook/mms\-lid\-4017, a language identification model which trained on and is capable of classifying 4,017 languages\. In contrast to the pure SSL model, MMS\-LID\-4017 is specifically optimized for language identification \(LID\), thereby emphasizing features that distinguish one language from another\. By comparing these two architectures, we evaluate two routing paradigms: SSL\-embedding\- and LID–embedding–based routing, respectively\.

Knowledge\-Guided Grouping\.Knowledge\-guided grouping constructs language clusters by leveraging established linguistic priors\. To provide a principled and quantitative representation of this knowledge, we utilizeURIEL\+language vectors\(Littellet al\.,[2017](https://arxiv.org/html/2607.24030#bib.bib54); Khanet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib55)\)\. These vectors integrate curated linguistic databases into language\-level representations, covering 7000\+ languages\. It also incorporates associated distance metrics derived from a multidimensional attribute space encompassing geographical proximity, typological features, morphological structures, and script\-related properties\.

Specifically, we construct language distance matrices using two distinct categories of URIEL\+ vectors:geographicandgenetic\. Geographic vectors capture spatial proximity between languages, capturing regional contact and shared acoustic environments\. In contrast, genetic vectors encode genealogical relationships through language family and subfamily hierarchies\. This design facilitates typology\-aware grouping grounded in shared linguistic ancestry and historical evolution rather than surface\-level feature similarity\. Consistent with the embedding\-guided approach, final language clusters are obtained by applying AHC with Ward’s linkage to the resulting distance matrices\.

### 3\.3Two\-Stage Dense\-to\-Sparse Upcycling

We apply a two\-stage upcycling strategy to the S3M\-based ASR models, inspired by LLM upcycling approaches that transform dense models into sparse MoE architectures\(Komatsuzakiet al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib71); Heet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib72); Leeet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib73)\)\. This approach allows us to preserve the rich representations learned by dense models while enabling efficient scaling, reducing training cost, and improving parameter utilization in sparse settings\. In the first stage, a dense ASR model is trained on top of pretrained S3M to establish robust global representations and transcription capabilities across languages\. In the second stage, we introduce the MoLGE framework to enable expert specialization across predefined language groups, upcycling the dense model into a sparse, parameter\-efficient architecture while further improving performance\. During the second stage, we apply LoRA to all linear layers, including attention projections and feed\-forward networks \(FFNs\)\. Motivated by prior work suggesting that FFNs encode model knowledge\(Yunet al\.,[2020](https://arxiv.org/html/2607.24030#bib.bib63); Gevaet al\.,[2021](https://arxiv.org/html/2607.24030#bib.bib64); Daiet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib65)\)while attention layers control information aggregation\(Abnar and Zuidema,[2020](https://arxiv.org/html/2607.24030#bib.bib66); Ferrandoet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib67)\), we adapt both components\. Restricting adaptation to attention layers alone is insufficient for modeling high\-density speech representations, making FFN adaptation essential for stable performance across diverse languages\.

## 4Experiments

### 4\.1Datasets

We combine three multilingual ASR datasets for training—FLEURS\(Conneauet al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib57)\)\(968 hours, 102 languages\), CommonVoice\(Ardilaet al\.,[2020](https://arxiv.org/html/2607.24030#bib.bib58)\)\(10,448 hours, 131 languages\), and the Omnilingual ASR Corpus\(Kerenet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib9)\)\(2,342 hours, 348 languages\)—totaling 13,758 hours across 495 unique languages\. All samples were resampled to 16 kHz\. As the Omnilingual ASR corpus mainly consists of long utterances, we segment them into approximately 30\-second chunks using MMS\-FA\(Pratapet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib7)\)\. The segmented OmniASR dataset is available at HuggingFace333https://huggingface\.co/datasets/Sanghyang00/omniasr\-molge, and additional details on preprocessing procedures are provided in Appendix[A](https://arxiv.org/html/2607.24030#A1)\.

### 4\.2Backbone and Baseline Models

We evaluate the MoLGE framework using two multilingual ASR models as a backbone, which emphasize complementary aspects of speech recognition:phoneticandorthographicmodeling\. For a fair comparison, both models utilize Connectionist Temporal Classification \(CTC\) variants for ASR and are configured with 300M parameters\. Details on the rationale behind the backbone selection are provided in Appendix[B](https://arxiv.org/html/2607.24030#A2)\.

LAMA\-UT\.LAMA\-UT\(Leeet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib56)\)is a multilingual ASR framework that predicts a universal Romanized transcription in a language\-agnostic manner, subsequently transliterating it into language\-specific scripts using frozen LLMs\. In our experiments, we use only the Romanized transcription generator to isolate the model’sphoneticmodeling capabilities\.

Omnilingual ASR\.Omnilingual ASR\(Kerenet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib9)\)is a massively multilingual ASR model based on wav2vec2\.0, supporting over 1600 languages\. In contrast to LAMA\-UT, this model directly predicts language\-specific scripts, emphasizingorthographicmodeling\.

Baseline Configurations\.For each backbone, we compare MoLGE against two baselines: \(1\) adense model, which uses a single ASR model without any adaptation, corresponding to the original CTC\-based ASR architecture, and \(2\)random grouping, where MoLGE is applied with randomly assigned language groups without any linguistic guidance\.

### 4\.3Training Details

We configured the MoLGE with a fixed LoRA rank of 32 and trained it with a composite objective of CTC loss\(Graves,[2012](https://arxiv.org/html/2607.24030#bib.bib59)\)and a routing loss, where implicit grouping uses load balancing and explicit grouping uses cross\-entropy for supervised group prediction\. The routing loss is weighted by 0\.01 relative to the CTC loss\. All models are trained with AdamW\(Loshchilov and Hutter,[2019](https://arxiv.org/html/2607.24030#bib.bib60)\)optimizer using a tri\-stage learning rate schedule over 10k steps: linear warm\-up from5​e−65\\mathrm\{e\}^\{\-6\}to1​e−41\\mathrm\{e\}^\{\-4\}\(10%\), constant \(60%\), and decay to5​e−65\\mathrm\{e\}^\{\-6\}\(30%\)\.

To alleviate resource and environmental imbalance in multilingual data, we adopt a multilevel sampling strategy followingPratapet al\.\([2024](https://arxiv.org/html/2607.24030#bib.bib7)\)andKerenet al\.\([2025](https://arxiv.org/html/2607.24030#bib.bib9)\), formulated aspl,d∝\(nl,d/N\)βL,Dp\_\{l,d\}\\propto\(n\_\{l,d\}/N\)^\{\\beta\_\{L,D\}\}\. Here,nl,dn\_\{l,d\}denotes the number of samples for languagell\(or corpusdd\), andNNis the total number of samples\. We setβL=βD=0\.5\\beta\_\{L\}=\\beta\_\{D\}=0\.5\. Additionally, we also apply data augmentation using room impulse responses \(RIR\)\(Szökeet al\.,[2019](https://arxiv.org/html/2607.24030#bib.bib61)\)and background noise\(Thiemannet al\.,[2013](https://arxiv.org/html/2607.24030#bib.bib62)\), where 50% of samples are clean, and the rest are equally split between RIR convolution and additive noise \(25% each\)\.

- •Low\-Resource:<<10h,Mid\-Resource: 10–50h,High\-Resource:\>\>50h

BackboneMethodActivatedParams\.AverageLow\-ResourceMid\-ResourceHigh\-ResourceCER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)CER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)CER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)CER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)LAMA\-UT\(Phonetic\)Dense300 M25\.15\-26\.16\-19\.06\-21\.84\-Random343 M24\.3168\.4425\.4669\.8817\.6677\.7719\.9323\.84MoLGE343 M22\.2793\.9623\.4394\.8415\.5593\.4317\.9980\.51OmniASR\(Orthographic\)Dense300 M29\.30\-31\.41\-20\.68\-13\.38\-Random353 M25\.8153\.2427\.5953\.3218\.1267\.8413\.3719\.37MoLGE353 M23\.7394\.4425\.4495\.2716\.3489\.3511\.7291\.91

Table 1:Performance on LAMA\-UT and Omnilingual ASR \(OmniASR\) models\. We report CER \(%\) and routing accuracy \(Acc\., %\) breakdown by resource availability\.

## 5Results and Analysis

![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/regression_unified.png)Figure 3:Regression plot between CER and training data\. Each point indicates the mean CER per language, and the shaded region represents the confidence interval\.### 5\.1Effects of MoLGE on Massively Multilingual ASR

Performance Across Linguistic Domains\.As shown in Table[1](https://arxiv.org/html/2607.24030#S4.T1), applying MoLGE yields substantial improvements in both phonetic and orthographic modeling for speech recognition, with relative gains of 11\.5% and 19\.0% in CER, respectively\. Interestingly, improvements in the phonetic domain are relatively modest compared to the orthographic domain, and mid\-resource languages sometimes exhibit lower CER than high\-resource languages\. This suggests that performance in the phonetic domain \(i\.e\., a simplified latent space\(Jaavidet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib68); Sajiet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib69); Liuet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib70)\)\) is influenced not only by data scale but also by factors such as representation complexity\. In contrast, in the orthographic domain, MoLGE significantly improves both CER and routing accuracy compared to dense and randomly grouped baselines, highlighting the benefits of incorporating language\-aware structure in more complex linguistic spaces\. Overall, these results indicate that the effectiveness of MoLGE is closely tied to linguistic complexity, with linguistic priors becoming increasingly beneficial for modeling richer orthographic and language\-specific structures\.

Data–Performance Correlation and Generalization Ability\.In Figure[3](https://arxiv.org/html/2607.24030#S5.F3), we illustrate the regression plot, which visualizes the correlation between CER and training data in the orthographic domain \(i\.e\., OmniASR\)\. When applying MoLGE, the correlation between training data size and CER becomes marginally stronger, with the correlation coefficientrrdecreasing from \-0\.24 to \-0\.26, leading to more predictable learning dynamics\. Moreover, the regression slope becomes less steep, with CER on low\-resource languages dropping substantially from 42% to 30%, demonstrating that MoLGE narrows the performance gap across resource levels\. Furthermore, confidence intervals become notably narrower, indicating that languages with similar training data now achieve more consistent performance\. Taken together, these results demonstrate that MoLGE enhances robustness in low\-resource languages and enables reliable generalization across languages of varying resource levels\.

![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/cluster_ssl.png)\(a\)SSL
![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/cluster_lid.png)\(b\)LID
![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/cluster_geographic.png)\(c\)Geographic
![Refer to caption](https://arxiv.org/html/2607.24030v1/figures/cluster_genetic.png)\(d\)Genetic

Figure 4:Visualization of language clustering across different grouping strategies\. Each point represents a language, and colors indicate cluster assignments in the t\-SNE space\.
### 5\.2Analysis on Cluster Formation and Language Grouping

Cluster Formation Across Guidance Methods\.Figure[4](https://arxiv.org/html/2607.24030#S5.F4)illustrates clustering results using t\-SNE across four explicit grouping strategies\. Embedding\-guided methods generally produce weakly separated clusters, with SSL representations showing particularly fragmented structures\. While LID\-based clustering yields clearer boundaries than SSL, it remains more dispersed than knowledge\-guided methods, likely due to the high\-dimensional nature of embeddings \(e\.g\., a feature dimensionality of 1,280 for both SSL and LID models\)\.

In contrast, knowledge\-guided methods produce more structured and well\-separated clusters than embedding\-guided approaches, owing to the compact nature of knowledge vectors\. Geographic grouping forms coarse but coherent clusters, while genetic grouping leaves some languages scattered near the center\. We attribute this behavior to the sensitivity of knowledge\-guided methods to the imputation mechanism of URIEL\+\. Since URIEL\+ imputes missing features based on information from related or higher\-level language families, genetic grouping often assigns linguistically diverse languages—including isolates such as Japanese \(jpn\) and Korean \(kor\), dialects such as Liberian English \(lir\), and Maria \(mrr\), and indigenous languages such as Hunjara\-Kaina Ke \(hkk\) and Ömie \(aom\)—to the same group\. Additional visualizations with language annotations are provided in Appendix[D](https://arxiv.org/html/2607.24030#A4)\.

LAMA\-UT \(Phonetic\)OmniASR \(Orthographic\)GuidanceMethodCER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)GuidanceMethodCER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)ImplicitFrame22\.30\-ImplicitFrame24\.30\-Sample22\.32\-Sample24\.51\-EmbeddingSSL22\.2972\.17EmbeddingSSL23\.8264\.16LID22\.2793\.96LID23\.7494\.70KnowledgeGeographic22\.2993\.51KnowledgeGeographic23\.7394\.44Genetic22\.2992\.24Genetic23\.8090\.79

Table 2:Average transcription performance \(CER\) and routing accuracy \(Acc\.\) on six guidance methods, respectively\.
- •S\.E\.: Shared Expert
- •H\.A\.: Hierarchical Adapt\.

ModelCER \(↓\\downarrow\)Acc\. \(↑\\uparrow\)MoLGE23\.7394\.44−\-S\. E\.24\.7092\.77−\-H\. A\. \(M\)23\.8573\.12−\-H\. A\. \(L\)27\.24\-

Table 3:Ablation study on architecture design\.

Effect of Cluster Coherence\.Table[3](https://arxiv.org/html/2607.24030#S5.T3)presents the performance of MoLGE under various grouping configurations\. Across both phonetic and orthographic spaces, explicit guidance consistently outperforms implicit guidance, supporting our hypothesis that prior\-based grouping is more effective than router\-driven grouping\.

To further analyze this trend, we break down the results by representation space\. In the phonetic space, performance differences across grouping strategies are minimal, also suggesting that language priors play a limited role in a simplified phonetic latent space\. In contrast, in the orthographic space, incorporating language priors consistently improves performance, and more coherent grouping leads to lower CER, as evidenced by the superior performance of LID over SSL embeddings and geographic over genetic vectors\.

Overall, these findings indicate that the effectiveness of language grouping arises from the interaction between linguistic complexity, the information density of linguistic priors, and cluster coherence\. As linguistic complexity increases, richer priors yield more coherent groupings, which in turn translate into improved performance\.

Indic LanguagesNigerian LanguagesLanguageHoursCER \(↓\\downarrow\)LanguageHoursCER \(↓\\downarrow\)Tamil89\.1314\.73Hausa12\.2412\.58Bengali40\.3810\.68Yoruba10\.8222\.46Marathi11\.449\.90Tangale7\.0120\.15Chitwania Tharu5\.4310\.70Burak6\.7019\.82Odia5\.0519\.01Tula5\.6821\.63Santali \(Ol Chiki\)0\.0721\.21Cakfem\-Mushere5\.5324\.00

Table 4:ASR performance of MoLGE on six Indic and Nigerian languages, respectively\.
\# ExpertsTotal Params\.CER \(↓\\downarrow\)8409M27\.3416465M24\.5132580M24\.52

Table 5:Effect of the number of experts in sample\-level implicit grouping\.

### 5\.3Knowledge Transfer Across Resource Levels

Table[5](https://arxiv.org/html/2607.24030#S5.T5)presents ASR performance on representative mid\- and low\-resource languages from India and Nigeria\. To investigate whether MoLGE promotes knowledge sharing across resource levels, we compare training hours and recognition performance within each language group\.

As shown in Table[5](https://arxiv.org/html/2607.24030#S5.T5), languages belonging to the same linguistic family are commonly assigned to the same group and exhibit relatively consistent CER trends despite large differences in training hours\. For example, several low\-resource languages achieve CERs comparable to those of substantially higher\-resource languages within the same group\. Similar patterns are observed across both the Indic and Nigerian language groups\. Overall, these findings suggest that MoLGE facilitates knowledge transfer within each language group, helping reduce performance disparities between high\- and low\-resource languages\.

### 5\.4Ablation Study

Architectural Design\.In Table[3](https://arxiv.org/html/2607.24030#S5.T3), we conduct an ablation study by sequentially removing key components, starting with the shared experts in the upper Transformer layer MoLE module, followed by the speech\-aware hierarchical adaptation\. Removing the shared experts degrades both CER and routing accuracy, highlighting their importance in effective acoustic modeling within the MoLE framework\.

Furthermore, eliminating the hierarchical adaptation layer consistently deteriorated model performance\. Replacing MoLGE with pure MoLE \(indicated by ’M’\) slightly impaired performance despite higher parameter counts, while replacing it with vanilla LoRA \(indicated by ’L’\) led to a severe performance drop\. This confirms the necessity of hierarchical adaptation for integrating language\-specific acoustic and textual features\.

Number of Experts\.In addition to the theoretical motivation discussed in Section[3\.2](https://arxiv.org/html/2607.24030#S3.SS2), we conducted a pilot study to examine the impact of the number of experts\. As shown in Table[5](https://arxiv.org/html/2607.24030#S5.T5), reducing the number of experts from 16 to 8 resulted in a substantial degradation in recognition performance\. Increasing the number of experts to 32, however, provided negligible improvement \(CER from 24\.51 to 24\.52\) while increasing the model size from 465M to 580M parameters\. Based on these empirical and theoretical bases, we adopted 16 experts, which offers the best trade\-off between recognition performance and computational efficiency\.

## 6Conclusion

In this paper, we introduced Mixture of Language Group Experts \(MoLGE\), an efficient and scalable framework for massively multilingual automatic speech recognition\. We show that incorporating architectural inductive biases on linguistic structure, together with language grouping, improves transcription performance while enabling scalable modeling across a large number of languages and mitigating the curse of multilinguality\. Specifically, our approach integrates speech\-oriented parameter\-efficient finetuning with structured language grouping, providing a unified framework for scalable massively multilingual speech recognition\. We further conduct a comprehensive analysis of diverse grouping strategies, demonstrating their effectiveness across varying resource conditions and their role in promoting expert specialization\. Finally, experiments across phonetic and orthographic domains reveal that the benefits of language\-aware modeling become more pronounced as linguistic complexity increases, with explicit language priors providing more coherent and effective guidance\. We hope this work serves as a foundation for scaling multilingual speech systems and inspires future research on structured, language\-aware modeling\.

## Acknowledgment

This work was supported by the National Research Foundation of Korea \(NRF\) grant funded by the Korea government Ministry of Science and ICT \(MSIT\) \(RS\-2026\-25468664\)\.

## References

- Quantifying attention flow in Transformers\.InProceedings of the Annual Meeting of the Association for Computational Linguistics,pp\. 4190–4197\.Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- J\. O\. Alabi, X\. Liu, D\. Klakow, and J\. Yamagishi \(2025\)AfriHuBERT: a self\-supervised speech representation model for African languages\.InInterspeech 2025,pp\. 4023–4027\.Cited by:[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- R\. Ardila, M\. Branson, K\. Davis, M\. Kohler, J\. Meyer, M\. Henretty, R\. Morais, L\. Saunders, F\. Tyers, and G\. Weber \(2020\)Common Voice: a massively\-multilingual speech corpus\.InProceedings of the Language Resources and Evaluation Conference,pp\. 4218–4222\.Cited by:[§4\.1](https://arxiv.org/html/2607.24030#S4.SS1.p1.1)\.
- A\. Babu, C\. Wang, A\. Tjandra, K\. Lakhotia, Q\. Xu, N\. Goyal, K\. Singh, P\. von Platen, Y\. Saraf, J\. Pino, A\. Baevski, A\. Conneau, and M\. Auli \(2022\)XLS\-R: self\-supervised cross\-lingual speech representation learning at scale\.InInterspeech 2022,pp\. 2278–2282\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- A\. Baevski, Y\. Zhou, A\. Mohamed, and M\. Auli \(2020\)Wav2vec 2\.0: a framework for self\-supervised learning of speech representations\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 12449–12460\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- R\. Bagat, I\. Illina, and E\. Vincent \(2025\)Mixture of LoRA experts for low\-resourced multi\-accent automatic speech recognition\.InInterspeech 2025,pp\. 1143–1147\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- M\. Bernard and H\. Titeux \(2021\)Phonemizer: text to phones transcription for multiple languages in python\.Journal of Open Source Software6\(68\),pp\. 3958\.Cited by:[§B\.2](https://arxiv.org/html/2607.24030#A2.SS2.p2.1)\.
- S\. Bharadwaj, C\. Li, Y\. Kim, K\. Choi, E\. Yeo, R\. S\. Shim, H\. Zhou, B\. Boldt, K\. R\. Jacome, K\. Chang,et al\.\(2026\)PRiSM: benchmarking phone realization in speech models\.arXiv preprint arXiv:2601\.14046\.Cited by:[§B\.2](https://arxiv.org/html/2607.24030#A2.SS2.p1.1)\.
- L\. Bloomfield \(1983\)An introduction to the study of language\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p5.1),[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p5.1)\.
- S\. Cahyawijaya \(2024\)LLM for everyone: Representing the underrepresented in large language models\.Hong Kong University of Science and Technology \(Hong Kong\)\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- S\. Chen, C\. Wang, Z\. Chen, Y\. Wu, S\. Liu, Z\. Chen, J\. Li, N\. Kanda, T\. Yoshioka, X\. Xiao,et al\.\(2022\)WavLM: large\-scale self\-supervised pre\-training for full stack speech processing\.IEEE Journal of Selected Topics in Signal Processing16\(6\),pp\. 1505–1518\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1)\.
- W\. Chen, J\. Tian, Y\. Peng, B\. Yan, C\. H\. Yang, and S\. Watanabe \(2025\)OWLS: scaling laws for multilingual speech recognition and translation models\.InInternational Conference on Machine Learning,Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- W\. Chen, W\. Zhang, Y\. Peng, X\. Li, J\. Tian, J\. Shi, X\. Chang, S\. Maiti, K\. Livescu, and S\. Watanabe \(2024\)Towards robust speech representation learning for thousands of languages\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,pp\. 10205–10224\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- A\. Conneau, A\. Baevski, R\. Collobert, A\. Mohamed, and M\. Auli \(2021\)Unsupervised cross\-lingual representation learning for speech recognition\.InInterspeech 2021,pp\. 2426–2430\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- A\. Conneau, K\. Khandelwal, N\. Goyal, V\. Chaudhary, G\. Wenzek, F\. Guzmán, E\. Grave, M\. Ott, L\. Zettlemoyer, and V\. Stoyanov \(2020\)Unsupervised cross\-lingual representation learning at scale\.InProceedings of the Annual Meeting of the Association for Computational Linguistics,pp\. 8440–8451\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- A\. Conneau, M\. Ma, S\. Khanuja, Y\. Zhang, V\. Axelrod, S\. Dalmia, J\. Riesa, C\. Rivera, and A\. Bapna \(2023\)FLEURS: few\-shot learning evaluation of universal representations of speech\.InIEEE Spoken Language Technology Workshop,pp\. 798–805\.Cited by:[§4\.1](https://arxiv.org/html/2607.24030#S4.SS1.p1.1)\.
- D\. Dai, L\. Dong, Y\. Hao, Z\. Sui, B\. Chang, and F\. Wei \(2022\)Knowledge neurons in pretrained Transformers\.InProceedings of the Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 8493–8502\.Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- D\. Fan, B\. Messmer, and M\. Jaggi \(2024\)Towards an empirical understanding of MoE design choices\.InICLR Workshop on Mathematical and Empirical Understanding of Foundation Models,Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p4.1)\.
- W\. Fedus, B\. Zoph, and N\. Shazeer \(2022\)Switch Transformers: scaling to trillion parameter models with simple and efficient sparsity\.Journal of Machine Learning Research23\(120\),pp\. 1–39\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p3.1)\.
- J\. Ferrando, G\. I\. Gállego, and M\. R\. Costa\-Jussà \(2022\)Measuring the mixing of contextual information in the Transformer\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,pp\. 8698–8714\.Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- A\. François \(2015\)Trees, waves and linkages: models of language diversification\.InThe Routledge Handbook of Historical Linguistics,pp\. 161–189\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p5.1)\.
- N\. Gaur, B\. Farris, P\. Haghani, I\. Leal, P\. J\. Moreno, M\. Prasad, B\. Ramabhadran, and Y\. Zhu \(2021\)Mixture of informed experts for multilingual speech recognition\.InProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing,pp\. 6234–6238\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p2.1)\.
- M\. Geva, R\. Schuster, J\. Berant, and O\. Levy \(2021\)Transformer feed\-forward layers are key\-value memories\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,pp\. 5484–5495\.Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- A\. Graves \(2012\)Connectionist temporal classification\.InSupervised Sequence Labelling with Recurrent Neural Networks,pp\. 61–93\.Cited by:[§4\.3](https://arxiv.org/html/2607.24030#S4.SS3.p1.3)\.
- D\. Gurgurov, T\. Bäumel, and T\. Anikina \(2024\)Multilingual large language models and curse of multilinguality\.arXiv preprint arXiv:2406\.10602\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- E\. He, A\. Khattar, R\. Prenger, V\. Korthikanti, Z\. Yan, T\. Liu, S\. Fan, A\. Aithal, M\. Shoeybi, and B\. Catanzaro \(2024\)Upcycling large language models into mixture of experts\.arXiv preprint arXiv:2410\.07524\.Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- P\. Heggarty, W\. Maguire, and A\. McMahon \(2010\)Splits or waves? Trees or webs? How divergence measures and network analysis can unravel language histories\.Philosophical Transactions of the Royal Society B: Biological Sciences365\(1559\),pp\. 3829–3843\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p5.1)\.
- N\. Houlsby, A\. Giurgiu, S\. Jastrzebski, B\. Morrone, Q\. De Laroussilhe, A\. Gesmundo, M\. Attariyan, and S\. Gelly \(2019\)Parameter\-efficient transfer learning for NLP\.InInternational Conference on Machine Learning,pp\. 2790–2799\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p1.1)\.
- W\. Hsu, B\. Bolte, Y\. H\. Tsai, K\. Lakhotia, R\. Salakhutdinov, and A\. Mohamed \(2021\)HuBERT: self\-supervised speech representation learning by masked prediction of hidden units\.IEEE/ACM Transactions on Audio, Speech, and Language Processing29,pp\. 3451–3460\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1)\.
- E\. J\. Hu, yelong shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p1.1)\.
- H\. Huang, S\. Lu, Y\. Shan, H\. Qu, F\. Zhang, W\. Guan, Q\. Hong, and L\. Li \(2025\)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\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- J\. Jaavid, R\. Dabre, M\. Aswanth, J\. Gala, T\. Jayakumar, R\. Puduppully, and A\. Kunchukuttan \(2024\)RomanSetu: efficiently unlocking multilingual capabilities of large language models via Romanization\.InProceedings of the Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 15593–15615\.Cited by:[§5\.1](https://arxiv.org/html/2607.24030#S5.SS1.p1.1)\.
- R\. A\. Jacobs, M\. I\. Jordan, S\. J\. Nowlan, and G\. E\. Hinton \(1991\)Adaptive mixtures of local experts\.Neural Computation3\(1\),pp\. 79–87\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p2.1)\.
- J\. M\. Janeiro, B\. Alastruey, F\. Massa, M\. Elbayad, B\. Piwowarski, P\. Gallinari, and L\. Barrault \(2025\)Mixture of languages: improved multilingual encoders through language grouping\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,pp\. 29695–29710\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- A\. Q\. Jiang, A\. Sablayrolles, A\. Roux, A\. Mensch, B\. Savary, C\. Bamford, D\. S\. Chaplot, D\. d\. l\. Casas, E\. B\. Hanna, F\. Bressand,et al\.\(2024\)Mixtral of experts\.arXiv preprint arXiv:2401\.04088\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p3.1)\.
- S\. B\. Kalluri, P\. Singh, P\. Roy Chowdhuri, A\. Kulkarni, S\. Baghel, P\. Hegde, S\. Sontakke, D\. K T, S\.R\. M\. Prasanna, D\. Vijayasenan, and S\. Ganapathy \(2024\)The Second DISPLACE Challenge: DIarization of SPeaker and LAnguage in Conversational Environments\.InInterspeech 2024,pp\. 1630–1634\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-1833),ISSN 2958\-1796Cited by:[Appendix C](https://arxiv.org/html/2607.24030#A3.p1.1)\.
- S\. Kalyan and A\. François \(2018\)Freeing the comparative method from the tree model: a framework for historical glottometry\.Senri Ethnological Studies98,pp\. 59–89\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p5.1)\.
- G\. Keren, A\. Kozhevnikov, Y\. Meng, C\. Ropers, M\. Setzler, S\. Wang, I\. Adebara, M\. Auli, C\. Balioglu,et al\.\(2025\)Omnilingual ASR: open\-source multilingual speech recognition for 1600\+ languages\.arXiv preprint arXiv:2511\.09690\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2607.24030#S4.SS1.p1.1),[§4\.2](https://arxiv.org/html/2607.24030#S4.SS2.p3.1),[§4\.3](https://arxiv.org/html/2607.24030#S4.SS3.p2.6)\.
- A\. Khan, M\. Shipton, D\. Anugraha, K\. Duan, P\. H\. Hoang, E\. Khiu, A\. S\. Doğruöz, and E\. A\. Lee \(2025\)URIEL\+: enhancing linguistic inclusion and usability in a typological and multilingual knowledge base\.InProceedings of the International Conference on Computational Linguistics,pp\. 6937–6952\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p7.1)\.
- A\. Komatsuzaki, J\. Puigcerver, J\. Lee\-Thorp, C\. R\. Ruiz, B\. Mustafa, J\. Ainslie, Y\. Tay, M\. Dehghani, and N\. Houlsby \(2023\)Sparse upcycling: training mixture\-of\-experts from dense checkpoints\.InInternational Conference on Learning Representations,Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- Y\. Kwon and S\. Chung \(2023\)MoLE: mixture of language experts for multi\-lingual automatic speech recognition\.InProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing,pp\. 1–5\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p2.1)\.
- J\. Lee, S\. Hwang, W\. Park, and M\. Ji \(2024\)Breaking ReLU barrier: generalized MoEfication for dense pretrained models\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- S\. Lee, W\. Chung, and H\. Kang \(2025\)LAMA\-UT: language agnostic multilingual ASR through orthography unification and language\-specific transliteration\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 24393–24401\.Cited by:[§4\.2](https://arxiv.org/html/2607.24030#S4.SS2.p2.1)\.
- D\. Lepikhin, H\. Lee, Y\. Xu, D\. Chen, O\. Firat, Y\. Huang, M\. Krikun, N\. Shazeer, and Z\. Chen \(2021\)GShard: scaling giant models with conditional computation and automatic sharding\.InInternational Conference on Learning Representations,Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p3.1)\.
- C\. Li, Y\. Deng, J\. Zhang, and C\. Zong \(2025a\)Group then scale: Dynamic mixture\-of\-experts multilingual language model\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 1730–1754\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- J\. Li, Y\. Shao, J\. Zhuo, C\. Li, L\. Tang, D\. Yu, and Y\. Qian \(2025b\)Efficient multilingual ASR finetuning via LoRA language experts\.InInterspeech 2025,pp\. 1138–1142\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- X\. Li, F\. Metze, D\. R\. Mortensen, S\. Watanabe, and A\. W\. Black \(2022\)Zero\-shot learning for grapheme to phoneme conversion with language ensemble\.InFindings of the Association for Computational Linguistics: ACL 2022,pp\. 2106–2115\.Cited by:[§B\.2](https://arxiv.org/html/2607.24030#A2.SS2.p2.1)\.
- Y\. Liang, F\. Meng, and J\. Zhou \(2025\)THOR\-MoE: hierarchical task\-guided and context\-responsive routing for neural machine translation\.InProceedings of the Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 21433–21445\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- P\. Littell, D\. R\. Mortensen, K\. Lin, K\. Kairis, C\. Turner, and L\. Levin \(2017\)URIEL and lang2vec: representing languages as typological, geographical, and phylogenetic vectors\.InProceedings of the Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers,pp\. 8–14\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p5.1),[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p7.1)\.
- Y\. Liu, M\. Wang, A\. H\. Kargaran, A\. ImaniGooghari, O\. Xhelili, H\. Ye, C\. Ma, F\. Yvon, and H\. Schütze \(2025\)How transliterations improve crosslingual alignment\.InProceedings of the International Conference on Computational Linguistics,pp\. 2417–2433\.Cited by:[§5\.1](https://arxiv.org/html/2607.24030#S5.SS1.p1.1)\.
- I\. Loshchilov and F\. Hutter \(2019\)Decoupled weight decay regularization\.InInternational Conference on Learning Representations,Cited by:[§4\.3](https://arxiv.org/html/2607.24030#S4.SS3.p1.3)\.
- L\. Lugo and V\. Vielzeuf \(2024\)Sustainable self\-supervised learning for speech representations\.arXiv preprint arXiv:2406\.07696\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p3.1)\.
- G\. Ma, W\. Wang, Y\. Li, Y\. Yang, B\. Du, and H\. Fu \(2023\)LAE\-ST\-MOE: boosted language\-aware encoder using speech translation auxiliary task for E2E code\-switching ASR\.InIEEE Automatic Speech Recognition and Understanding Workshop,pp\. 1–8\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- G\. W\. Milligan and M\. C\. Cooper \(1985\)An examination of procedures for determining the number of clusters in a data set\.Psychometrika50\(2\),pp\. 159–179\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p2.4)\.
- B\. Mu, K\. Wei, Q\. Shao, Y\. Xu, and L\. Xie \(2025\)HDMoLE: mixture of LoRA experts with hierarchical routing and dynamic thresholds for fine\-tuning LLM\-based ASR models\.InProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing,pp\. 1–5\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p2.1)\.
- T\. Niesleret al\.\(2018\)A first south african corpus of multilingual code\-switched soap opera speech\.InProceedings of the eleventh international conference on language resources and evaluation \(LREC 2018\),Cited by:[Appendix C](https://arxiv.org/html/2607.24030#A3.p1.1)\.
- K\. Okabe, T\. Koshinaka, and K\. Shinoda \(2018\)Attentive statistics pooling for deep speaker embedding\.InInterspeech 2018,pp\. 2252–2256\.Cited by:[§3\.1](https://arxiv.org/html/2607.24030#S3.SS1.p3.1)\.
- A\. Pasad, J\. Chou, and K\. Livescu \(2021\)Layer\-wise analysis of a self\-supervised speech representation model\.InIEEE Automatic Speech Recognition and Understanding Workshop,pp\. 914–921\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p7.1),[§3\.1](https://arxiv.org/html/2607.24030#S3.SS1.p2.1)\.
- A\. Pasad, B\. Shi, and K\. Livescu \(2023\)Comparative layer\-wise analysis of self\-supervised speech models\.InProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing,pp\. 1–5\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p7.1),[§3\.1](https://arxiv.org/html/2607.24030#S3.SS1.p2.1)\.
- J\. N\. Pava, C\. Meinhardt, H\. B\. U\. Zaman, T\. Friedman, S\. T\. Truong, D\. Zhang, V\. Marivate, and S\. Koyejo \(2025\)Mind the \(language\) gap: Mapping the challenges of LLM development in low\-resource language contexts\.Stanford Institute for Human\-Centered Artificial Intelligence \(HAI\): Stanford, CA, USA\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- Y\. Peng, M\. Shakeel, Y\. Sudo, W\. Chen, J\. Tian, C\. Lin, and S\. Watanabe \(2025\)OWSM v4: improving open Whisper\-style speech models via data scaling and cleaning\.InInterspeech 2025,pp\. 2225–2229\.Cited by:[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- J\. Pfeiffer, A\. Kamath, A\. Rücklé, K\. Cho, and I\. Gurevych \(2021\)AdapterFusion: non\-destructive task composition for transfer learning\.InProceedings of the Conference of the European Chapter of the Association for Computational Linguistics: Main Volume,pp\. 487–503\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p1.1)\.
- V\. Pratap, A\. Tjandra, B\. Shi, P\. Tomasello, A\. Babu, S\. Kundu, A\. Elkahky, Z\. Ni, A\. Vyas, M\. Fazel\-Zarandi,et al\.\(2024\)Scaling speech technology to 1,000\+ languages\.Journal of Machine Learning Research25\(97\),pp\. 1–52\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1),[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2607.24030#S4.SS1.p1.1),[§4\.3](https://arxiv.org/html/2607.24030#S4.SS3.p2.6)\.
- A\. Radford, J\. W\. Kim, T\. Xu, G\. Brockman, C\. McLeavey, and I\. Sutskever \(2023\)Robust speech recognition via large\-scale weak supervision\.InInternational Conference on Machine Learning,pp\. 28492–28518\.Cited by:[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- A\. Saji, J\. A\. Husain, T\. Jayakumar, R\. Dabre, A\. Kunchukuttan, and R\. Puduppully \(2025\)RomanLens: the role of latent Romanization in multilinguality in LLMs\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 26410–26429\.Cited by:[§5\.1](https://arxiv.org/html/2607.24030#S5.SS1.p1.1)\.
- J\. Shi, D\. Berrebbi, W\. Chen, E\. Hu, W\. Huang, H\. Chung, X\. Chang, S\. Li, A\. Mohamed, H\. Lee, and S\. Watanabe \(2023\)ML\-SUPERB: Multilingual Speech Universal PERformance Benchmark\.InInterspeech 2023,pp\. 884–888\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2023-1316),ISSN 2958\-1796Cited by:[§B\.1](https://arxiv.org/html/2607.24030#A2.SS1.p1.1)\.
- J\. Shi, S\. Wang, W\. Chen, M\. Bartelds, V\. Bannihatti Kumar, J\. Tian, X\. Chang, D\. Jurafsky, K\. Livescu, H\. Lee, and S\. Watanabe \(2024\)ML\-SUPERB 2\.0: Benchmarking Multilingual Speech Models Across Modeling Constraints, Languages, and Datasets\.InInterspeech 2024,pp\. 1230–1234\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-2248),ISSN 2958\-1796Cited by:[§B\.1](https://arxiv.org/html/2607.24030#A2.SS1.p1.1)\.
- I\. Szöke, M\. Skácel, L\. Mošner, J\. Paliesek, and J\. Černockỳ \(2019\)Building and evaluation of a real room impulse response dataset\.IEEE Journal of Selected Topics in Signal Processing13\(4\),pp\. 863–876\.Cited by:[§4\.3](https://arxiv.org/html/2607.24030#S4.SS3.p2.6)\.
- J\. Thiemann, N\. Ito, and E\. Vincent \(2013\)The diverse environments multi\-channel acoustic noise database \(DEMAND\): a database of multichannel environmental noise recordings\.InProceedings of Meetings on Acoustics,Vol\.19,pp\. 035081\.Cited by:[§4\.3](https://arxiv.org/html/2607.24030#S4.SS3.p2.6)\.
- W\. Wang, G\. Ma, Y\. Li, and B\. Du \(2023\)Language\-routing mixture of experts for multilingual and code\-switching speech recognition\.InInterspeech 2023,pp\. 1389–1393\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- Z\. Wang, Z\. C\. Lipton, and Y\. Tsvetkov \(2020\)On negative interference in multilingual models: Findings and a meta\-learning treatment\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,pp\. 4438–4450\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- J\. H\. Ward Jr \(1963\)Hierarchical grouping to optimize an objective function\.Journal of the American Statistical Association58\(301\),pp\. 236–244\.Cited by:[§3\.2](https://arxiv.org/html/2607.24030#S3.SS2.p5.1)\.
- X\. Wu, S\. Huang, and F\. Wei \(2024\)Mixture of LoRA experts\.InInternational Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p2.1)\.
- Y\. Yang, J\. Zhuo, Z\. Jin, Z\. Ma, X\. Yang, Z\. Yao, L\. Guo, W\. Kang, F\. Kuang, L\. Lin,et al\.\(2025\)k2SSL: a faster and better framework for self\-supervised speech representation learning\.InIEEE International Conference on Multimedia and Expo,pp\. 1–6\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p3.1)\.
- C\. Yun, S\. Bhojanapalli, A\. S\. Rawat, S\. Reddi, and S\. Kumar \(2020\)Are Transformers universal approximators of sequence\-to\-sequence functions?\.InInternational Conference on Learning Representations,Cited by:[§3\.3](https://arxiv.org/html/2607.24030#S3.SS3.p1.1)\.
- M\. Zanon Boito, V\. Iyer, N\. Lagos, L\. Besacier, and I\. Calapodescu \(2024\)mHuBERT\-147: a compact multilingual HuBERT model\.InInterspeech 2024,pp\. 3939–3943\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.24030#S2.SS1.p1.1)\.
- S\. Zhang, T\. Parcollet, R\. van Dalen, and S\. Bhattacharya \(2024\)Linear\-complexity self\-supervised learning for speech processing\.InInterspeech 2024,pp\. 3480–3484\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p3.1)\.
- Y\. Zhang, W\. Han, J\. Qin, Y\. Wang, A\. Bapna, Z\. Chen, N\. Chen, B\. Li, V\. Axelrod, G\. Wang,et al\.\(2023\)Google USM: scaling automatic speech recognition beyond 100 languages\.arXiv preprint arXiv:2303\.01037\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p2.1)\.
- G\. Zheng, X\. Wang, J\. Liang, N\. Chen, Y\. Zheng, and B\. Wang \(2024\)Efficiently democratizing medical LLMs for 50 languages via a mixture of language family experts\.arXiv preprint arXiv:2410\.10626\.Cited by:[§2\.2](https://arxiv.org/html/2607.24030#S2.SS2.p3.1)\.
- T\. Zhong, Z\. Yang, Z\. Liu, R\. Zhang, W\. You, Y\. Liu, H\. Sun, Y\. Pan, Y\. Li, Y\. Zhou,et al\.\(2024\)Opportunities and challenges of large language models for low\-resource languages in humanities research\.arXiv preprint arXiv:2412\.04497\.Cited by:[§1](https://arxiv.org/html/2607.24030#S1.p4.1)\.
- J\. Zhu, F\. Samir, E\. Chodroff, and D\. R\. Mortensen \(2025\)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\.Cited by:[§B\.2](https://arxiv.org/html/2607.24030#A2.SS2.p1.1)\.
- J\. Zhu, C\. Zhang, and D\. Jurgens \(2022\)ByT5 model for massively multilingual grapheme\-to\-phoneme conversion\.External Links:[Link](https://arxiv.org/abs/2204.03067),[Document](https://dx.doi.org/10.48550/ARXIV.2204.03067)Cited by:[§B\.2](https://arxiv.org/html/2607.24030#A2.SS2.p2.1)\.

![Refer to caption](https://arxiv.org/html/2607.24030v1/figures_appendix/duration_file.png)\(a\)Duration of the original dataset\.
![Refer to caption](https://arxiv.org/html/2607.24030v1/figures_appendix/duration_segment.png)\(b\)Duration of the dataset after segmentation\.

Figure 5:Duration distributions in the omnilingual ASR corpus, aggregated across all splits\.## Appendix ADataset Details and Preprocessing Strategy

### A\.1Dataset Details

FLEURS\.FLEURS is a multilingual dataset encompassing 102 languages, specifically curated for ASR and speech translation\. The corpus consists of read speech from Wikipedia articles, providing relatively clean recordings\. Each language contains approximately 10 hours of supervised data, offering a balanced foundation for high\-quality acoustic modeling\.

CommonVoice\.CommonVoice is a large\-scale crowdsourced dataset covering 131 languages, exhibiting substantial acoustic variability, including diverse microphone qualities and background noise levels\. We use the official splits from CommonVoice 22\.0\.

Omnilingual ASR Corpus\.Omnilingual ASR corpus is a massively multilingual dataset spanning 348 languages, with a significant emphasis on low\-resource languages and long\-form speech\. To mitigate training instability associated with long utterances, we perform audio segmentation via forced alignment\.

### A\.2Preprocessing Details of Omnilingual ASR Corpus

Dataset Characteristics\.The Omnilingual ASR corpus differs from conventional ASR datasets in that it captures naturally occurring speech by prompting speakers with over 1,500 queries and recording their responses\. As a result, it contains many long utterances exceeding 30 seconds, as shown in Figure[5](https://arxiv.org/html/2607.24030#A0.F5)\. 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 \(e\.g\., periods, question marks, exclamation marks\), \(2\) segmentation using soft punctuation \(e\.g\., commas\), and \(3\) forced segmentation based on a character limit\. While most utterances can be segmented within 30 seconds using punctuation, we observe that some languages—particularly African languages—often contain long, punctuation\-free running sentences\. To address this, we apply forced segmentation at approximately 150 characters, splitting at the nearest word boundary\. All segmentation is performed using MMS\-FA\. After segmentation, we filter samples based on alignment scores on a per\-language basis, pruning low\-quality segments to ensure overall data quality\.

## Appendix BUtilization of Romanization as a Phonetic Proxy

### B\.1Representation Analysis in Multilingual ASR

Recent multilingual speech benchmarks, such as ML\-SUPERB\(Shiet al\.,[2023](https://arxiv.org/html/2607.24030#bib.bib74);[2024](https://arxiv.org/html/2607.24030#bib.bib75)\), have demonstrated the effectiveness of self\-supervised speech models for multilingual ASR across diverse languages\. Most existing studies, however, primarily focus on improving downstream recognition performance through larger pretrained models, scaling strategies, or multilingual evaluation benchmarks\.

In contrast, relatively little attention has been paid to how different transcription representations influence multilingual speech modeling\. In particular, it remains unclear whether language\-aware modeling techniques exhibit different behaviors when operating on orthographic or phonetic representations\. Since our goal is not to introduce a new ASR backbone but to understand the effectiveness of language\-aware expert grouping, we intentionally analyze MoLGE under two complementary representation spaces\.

### B\.2Romanization as a Scalable Alternative

Ideally, phonetic representations would be constructed using language\-specific phoneme sequences, such as IPA\-based transcriptions\. In this context, we also considered adopting multilingual phonetic ASR benchmarks, such as ZIPA\(Zhuet al\.,[2025](https://arxiv.org/html/2607.24030#bib.bib76)\)and PRiSM\(Bharadwajet al\.,[2026](https://arxiv.org/html/2607.24030#bib.bib77)\)\. 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\. Existing multilingual G2P toolkits provide only limited language coverage\. For example, Phonemizer\(Bernard and Titeux,[2021](https://arxiv.org/html/2607.24030#bib.bib78)\)and CharsiuG2P\(Zhuet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib79)\)support approximately 120 and 100 languages, respectively, while broader\-coverage systems such as Transphone\(Liet al\.,[2022](https://arxiv.org/html/2607.24030#bib.bib80)\)rely extensively on zero\-shot G2P generation, making phonetic quality difficult 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 broader range of languages while preserving meaningful phonetic information\. Consequently, it has been adopted in several large\-scale multilingual speech systems, including MMS, where Romanized transcriptions are used for nearly one thousand languages\. Although Romanization is not equivalent to true phonetic representations, prior studies have demonstrated that it serves as an effective phonetic proxy for multilingual speech modeling\.

## Appendix CCode\-Switching and MoLGE

To further investigate the linguistic behavior of MoLGE under code\-switching scenarios, we analyze two public corpora with time\-aligned language annotations: DISPLACE 2024\(Kalluriet al\.,[2024](https://arxiv.org/html/2607.24030#bib.bib81)\)and the South African Soap Opera corpus\(Niesler and others,[2018](https://arxiv.org/html/2607.24030#bib.bib82)\)The analysis indicates that, under sentence\-level routing, the assigned expert group is largely determined 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\.

This behavior is likely attributable to the receptive field of the S3M backbone, where each frame represents approximately 25ms of acoustic context\. Such short temporal windows are primarily optimized for capturing local phonetic information, making it difficult to reliably infer language identity at the frame level, particularly in code\-switching speech\.

## Appendix DDetailed Visualization on Language Clusters

![Refer to caption](https://arxiv.org/html/2607.24030v1/figures_appendix/cluster_ssl_app.png)\(a\)Cluster formation using the SSL embedding\.
![Refer to caption](https://arxiv.org/html/2607.24030v1/figures_appendix/cluster_lid_app.png)\(b\)Cluster formation using the LID embedding\.

Figure 6:Comparison of embedding\-guided clusters\.![Refer to caption](https://arxiv.org/html/2607.24030v1/figures_appendix/cluster_geographic_app.png)\(a\)Cluster formation using the geographic vector\.
![Refer to caption](https://arxiv.org/html/2607.24030v1/figures_appendix/cluster_genetic_app.png)\(b\)Cluster formation using the genetic vector\.

Figure 7:Comparison of knowledge\-guided clusters\.In this section, we provide detailed visualizations of the clustering results for each explicit grouping strategy, as illustrated in Figures[6](https://arxiv.org/html/2607.24030#A4.F6)and[7](https://arxiv.org/html/2607.24030#A4.F7)\. Languages assigned to the same group are indicated by the same color\. Each plot is annotated with ISO 639\-3 language codes following the Ethnologue444https://www\.ethnologue\.com/convention\.