ELARSpeech：涵盖675种代表性不足语言的3万小时数据集，面向长尾多语言语音技术

# ELARSpeech: A 30K-Hour Dataset Across 675 Underrepresented Languages for Long-Tail Multilingual Speech Technologies
Source: [https://arxiv.org/html/2610.19388](https://arxiv.org/html/2610.19388)
Yejin JeonAffiliation:McGill UniversityAffiliation:Mila \- Quebec AI InstituteDavid Ifeoluwa AdelaniAffiliation:McGill UniversityAffiliation:Mila \- Quebec AI InstituteAffiliation:Canada CIFAR AI ChairCorrespondence:[sangmin\_lee@dsp\.yonsei\.ac\.kr,david\.adelani@mila\.quebec](mailto:[email protected],[email protected])

###### Abstract

Massively multilingual speech modeling has made remarkable progress in recent years, yet existing training resources and benchmarks still underrepresent much of the linguistic long tail\. To address this gap, we introduce ELARSpeech, a large\-scale multilingual speech resource comprising 30,202 hours of speech across 675 underrepresented languages, curated from the Endangered Languages Archive\. Combining human\-annotated and unlabeled recordings, the corpus spans diverse recording conditions, domains, and language families\. Beyond corpus construction, our experiments show how ELARSpeech supports extending ASR and TTS capabilities to diverse underrepresented languages, with further benefits from pseudo\-labeling unlabeled data\. At the same time, our speech representation analysis reveals remaining limitations in transferring existing multilingual models to the linguistic long tail\. Together, ELARSpeech provides both a resource and a testbed for extending multilingual speech technology further into the linguistic long tail\. We publicly release the dataset, all trained ASR and TTS models, and the complete data construction pipeline to facilitate research on long\-tail multilingual speech technologies\.111https://github\.com/sanghyang00/elarspeech/

## 1Introduction

Recent advances in multilingual speech modeling have dramatically expanded the linguistic coverage of speech technologies, enabling automatic speech recognition[Radford et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib1);[Zhang et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib2);[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3);[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4)and speech synthesis[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3);[Zhu et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib5)across hundreds to thousands of languages\. As modeling approaches become more sophisticated and models continue to scale successfully, continued progress toward the linguistic long tail requires multilingual speech resources to advance in both scale and diversity\.

Meanwhile, parallel efforts have significantly increased the volume and language coverage of multilingual speech resources\. Large\-scale efforts such as MMS[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3)provide labeled speech for around 1,000 languages alongside unlabeled speech for around 3,000 languages, and recent web\-crawled corpora[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13)have further expanded the diversity of resources\. While these advances have considerably expanded multilingual speech resources, three key challenges remain in supporting the linguistic long tail\.

The first challenge is the gap between language coverage and accessibility\. Publicly available corpora typically cover only a few hundred languages[Ardila et al\. \(2020\)](https://arxiv.org/html/2610.19388#bib.bib6);[Valk and Alumäe \(2021\)](https://arxiv.org/html/2610.19388#bib.bib7);[Conneau et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib8);[Li et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib9), whereas datasets spanning thousands of languages[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3);[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4)often follow restrictive licenses or are not publicly released\. Consequently, speech data for many long\-tail languages remain difficult to access, limiting reproducible research further into the linguistic long tail\.

The second challenge is constrained acoustic and domain diversity\. Existing multilingual corpora draw heavily from a single source, such as parliamentary proceedings[Wang et al\. \(2021\)](https://arxiv.org/html/2610.19388#bib.bib10);[Pfisterer et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib11);[Asonitis et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib12)or Bible readings[Meyer et al\. \(2022\)](https://arxiv.org/html/2610.19388#bib.bib39);[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13);[Guzmán et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib14), which naturally concentrate the resulting data around particular recording conditions, speaking styles, and domains\. This leaves broader communicative settings, including spontaneous and conversational speech recorded in diverse real\-world environments, less represented\.

The third challenge is the lack of suitable evaluation resources for the linguistic long tail\. Despite advances in multilingual speech foundation models trained on increasingly large and diverse language collections[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3);[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13);[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4), their effectiveness on underrepresented languages remains insufficiently characterized\. Existing multilingual evaluations[Shi et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib43);[Shi et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib44)tend to focus on higher\-resource languages, making it challenging to systematically assess whether improvements in multilingual pretraining translate into stronger representations for long\-tail languages[Boulianne \(2022\)](https://arxiv.org/html/2610.19388#bib.bib40);[Le Ferrand et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib41);[Li and Niehues \(2026\)](https://arxiv.org/html/2610.19388#bib.bib42)\.

To address these challenges, we introduceELARSpeech, a massively multilingual speech corpus that extends speech resource coverage into the linguistic long tail\. By drawing on field recordings from diverse languages, it enables training and evaluation of speech models on broader underrepresented languages\. In terms of scale,ELARSpeech comprises 30,202 hours across 675 underrepresented languagesand language varieties \(hereafter, languages\), making it one of the largest publicly available speech resources for underrepresented languages\. This includes 2,791 hours of labeled speech\-text pairs across 443 languages and 27,412 hours of unlabeled speech across 665 languages\.

In experiments, we investigate automatic speech recognition \(ASR\) and text\-to\-speech \(TTS\) adaptation, pseudo\-labeled scaling, and representation analysis using ELARSpeech\. Our ASR and TTS experiments demonstrate that ELARSpeech supports effective adaptation across diverse underrepresented languages, while pseudo\-labeling further improves performance by leveraging unlabeled speech\. Additionally, we evaluate existing speech self\-supervised models \(S3Ms\) on ELARSpeech, showing that advances in conventional multilingual settings do not necessarily transfer to the linguistic long tail\. These findings highlight the broader utility of ELARSpeech as a resource for advancing speech technologies in underrepresented languages and a foundation for evaluating speech models across the linguistic long tail\.

Our contributions are summarized as follows:

- •We introduce ELARSpeech, a massively multilingual speech corpus comprising 30,202 hours of curated speech from 675 languages, encompassing both human\-labeled and unlabeled subsets\.
- •ELARSpeech substantially extends the long tail of open multilingual speech resources while preserving the linguistic, acoustic, and domain diversity of real\-world field recordings\.
- •We demonstrate ELARSpeech’s utility in extending ASR and TTS capabilities to underrepresented languages and uncover challenges in transferring existing S3Ms to the linguistic long tail\.

## 2Related Work

### 2\.1Labeled Multilingual Speech Corpora

Labeled speech corpora have advanced considerably in both language coverage and scale\. Early efforts such as CommonVoice[Ardila et al\. \(2020\)](https://arxiv.org/html/2610.19388#bib.bib6)and FLEURS[Conneau et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib8)provide read speech through crowdsourcing and controlled recordings, respectively, while the more recent WorldSpeech[Asonitis et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib12)scales transcribed speech to over 65K hours\. For underrepresented languages, the Omnilingual ASR corpus[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4)extends ASR data to 348 languages, while OpenBibleTTS[Guzmán et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib14)provides TTS data for 37 languages derived from Bible recordings\. Nevertheless, corpora combining broad low\-resource language coverage with diverse environments still remain scarce\.

DatasetLabeledContent\# Langs\.Size \(H\)Closed\-Source DatasetsGoogle USM×\\timesVariety56812\.1MMMS\-UnlabBible28098KWhisper✓\\checkmarkVariety98680KMMS\-LabBible110745KALLASRVariety1690121KOpen\-Source DatasetsMMS\-Unlab v2×\\timesBible2778838WikitonguesConversation70070JesusDramasBible430643VoxLingua107Variety1077KWAXAL△\\triangleSpontaneous3216KYODAS v3Variety1471\.1MFLEURS✓\\checkmarkWikipedia1021KCommonVoice26Wikipedia29014KOpenBibleTTSBible424KWorldSpeechParliament7665KOmnilingual ASRSpontaneous3483KOursELARSpeech△\\triangleVariety67530K

Table 1:Comparison of existing low\-resource multilingual speech datasets\. "△\\triangle" indicates the mixture across labeled, pseudo\-labeled, and unlabeled data\.
### 2\.2Unlabeled Multilingual Speech Corpora

Unlabeled speech resources enable collections at even broader linguistic scales; MMS[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3)collected Bible recordings across thousands of languages, though its underlying corpus is not publicly distributed\. Subsequently, XEUS[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13)further improves accessibility by reconstructing the unlabeled MMS collection while crawling additional long\-tail speech resources from JesusDramas and WikiTongues\. Despite their broad language coverage, these resources still leave room to further diversify recording domains and conditions with real\-world speech data\.

\(a\)Global language coverage of ELARSpeech in the Mollweide projection[Snyder \(1997\)](https://arxiv.org/html/2610.19388#bib.bib26)\.30,202 H11 domainsRead 0\.20%Conversational 8\.92%Narrative 11\.82%Interview 3\.60%Elicited 16\.80%Ceremonial 3\.05%Song / Chant 1\.53%Lecture / Speech 0\.38%Broadcast / Media 0\.04%Mixed 5\.10%Others 48\.55%

\(b\)Domain distribution of ELARSpeech\.
30,202 H78 familiesAustronesian 15\.08%Niger\-Congo 13\.99%Sino\-Tibetan 10\.70%Others 8\.03%Afro\-Asiatic 5\.18%Trans\-New Guinea 3\.73%Language isolate 3\.57%Otomanguean 2\.97%Tupian 2\.33%Australian 2\.32%Mayan 2\.25%Uralic 2\.18%Rest \(67 families\) 27\.67%

\(c\)Language family distribution of ELARSpeech\.

Figure 1:Geographic, domain, and language family distributions of ELARSpeech\.

## 3ELARSpeech

### 3\.1Corpus Overview

#### Design Objectives\.

We design ELARSpeech with two main objectives: \(1\) extending resource coverage toward underrepresented languages that remain unsupported by existing multilingual speech models, and \(2\) broadening acoustic and domain diversity across diverse communicative settings and recording environments\.

#### Data Source\.

To meet these objectives, we select the Endangered Languages Archive222https://www\.elararchive\.org/\(ELAR\) as our primary data source\. Specifically, ELAR is a digital archive dedicated to preserving endangered languages through multimodal language documentation, hosting diverse resources including audio and video recordings, transcriptions, and field notes contributed by language communities and field linguists\. However, as ELAR is designed primarily for language documentation and preservation rather than as a training dataset, its recordings are largely uncurated, consisting of long\-form audio with heterogeneous recording conditions and varying levels of transcription availability\. We therefore develop a curation pipeline to transform these archival collections into a standardized speech dataset, as detailed in Sections[3\.2](https://arxiv.org/html/2610.19388#S3.SS2)and[3\.3](https://arxiv.org/html/2610.19388#S3.SS3)\. Following ELAR’s licensing requirements, ELARSpeech is distributed under the CC BY\-NC\-SA 4\.0 license, with details on the access permissions and licensing conditions of the underlying collections provided in Appendix[A](https://arxiv.org/html/2610.19388#A1)\.

#### Corpus Characteristics\.

ELARSpeech comprises 30,202 hours of speech spanning 675 languages, with their geographic, language\-family, and domain distributions summarized in Figure[1](https://arxiv.org/html/2610.19388#S2.F1)\. The corpus exhibits broad linguistic diversity while dramatically extending speech resource coverage toward the long tail\. Geographically, ELARSpeech covers languages distributed across diverse continents \(Figure[1\(a\)](https://arxiv.org/html/2610.19388#S2.F1.sf1)\), while its linguistic coverage spans 78 language families \(Figure[1\(c\)](https://arxiv.org/html/2610.19388#S2.F1.sf3)\)\. Among the 675 languages, 259 are not supported by existing multilingual models trained on public[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13)or non\-public resources[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3);[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4), and 53 are not represented in the ISO 639\-3 standard\. Beyond linguistic coverage, ELARSpeech captures broad diversity in communicative domains and recording conditions inherited from archival language documentation\. As shown in Figure[1\(b\)](https://arxiv.org/html/2610.19388#S2.F1.sf2), the corpus encompasses conversations, narratives, interviews, elicitation sessions, and read speech, with considerable variation in recording environments\. Detailed curation principles and additional corpus statistics are provided in Appendices[B](https://arxiv.org/html/2610.19388#A2)and[1](https://arxiv.org/html/2610.19388#alg1), respectively\.

#### Corpus Composition\.

As summarized in Table[2](https://arxiv.org/html/2610.19388#S3.T2), ELARSpeech consists of two subsets: \(1\)ELARSpeech\-L, containing 2,791 hours of human\-annotated speech across 443 languages, and \(2\)ELARSpeech\-U, containing 27,412 hours of preprocessed unlabeled speech across 665 languages\. For both subsets, recordings associated with multiple languages or those that cannot be reliably assigned to a single language are categorized as a separate multilingual split\. Subsequently, ELARSpeech\-L is split at the session level into training, development, and test sets with an 85/5/10 ratio, while languages with less than 5 minutes of development data are retained only in the training set\. Two subsets are constructed through separate curation pipelines as described below\.

TypeSplit\# Lang\.\# Utt\. \(M\)Duration \(H\)LabeledTrain4093\.092367\.08Dev2330\.17132\.77Test2330\.35266\.14Multi820\.0424\.82Total4433\.642790\.81UnlabeledTrain59219\.1426763\.37Multi2000\.41648\.29Total66519\.5627411\.65Overall67523\.2030202\.46

Table 2:Dataset statistics of ELARSpeech\. “Multi” denotes recordings categorized as multilingual\.

### 3\.2Constructing Labeled Subset

When timestamped annotations are available, we segment recordings into utterances using timestamps and pair them with annotations to construct ELARSpeech\-L through the following three stages\.

#### Audio\-Annotation Matching

We first identify candidate audio\-annotation pairs using filename matching\. In most ELAR collections, transcription files are named after their corresponding audio recordings, occasionally with additional metadata such as the annotation date appended to the filename \(e\.g\.,AUD001\_20260101\.eaf\)\. Accordingly, we first normalize all filenames to lowercase and then pair an audio file with a transcription file whenever the annotation filename contains the audio filename as an exact substring\.

#### Valid Annotation Parsing\.

The most challenging step in constructing ELARSpeech\-L is identifying valid transcription tiers, as annotation tier names are not standardized across ELAR collections and vary substantially across field linguists\. For example, transcriptions may appear in a dedicated tier such astxor in tiers named after individual speakers\. We therefore develop a tier selection algorithm that scores candidate tiers based on common naming conventions\. High\-confidence tiers are directly accepted, medium\-confidence tiers undergo manual verification, and low\-confidence tiers are discarded\. Detailed scoring criteria, thresholds, and the algorithm are provided in Appendix[D](https://arxiv.org/html/2610.19388#A4)\.

#### Speech Segmentation\.

After selecting valid transcription tiers, we segment each recording using the timestamps from the accepted tiers to obtain time\-aligned speech\-text pairs\. When annotation timestamps fall outside the bounds of the corresponding recording, we interpret this as potential audio\-annotation misalignment and discard all segments derived from the affected recording\.

### 3\.3Constructing Unlabeled Subset

For recordings without timestamped annotations, we preprocess the long\-form audio through the following pipeline to construct ELARSpeech\-U\.

#### Overlapping Recording Exclusion\.

First of all, we exclude all source recordings included in ELARSpeech\-L from the unlabeled pool to ensure that the two subsets remain disjoint, preventing potential data leakage in downstream experiments\.

#### Quality\-Based Pruning\.

Subsequently, we assess the acoustic quality of each recording usingBrouhaha[Lavechin et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib17), which estimates the signal\-to\-noise ratio \(SNR\) and speech clarity \(C50\)\. Recordings with either SNR or C50 below 0 dB are discarded to remove severely degraded audio\. Unlike previous in\-the\-wild dataset construction pipelines[He et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib15);[Ma et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib27);[Ogun et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib28)that apply speech enhancement, we use these quality estimates solely for filtering[Kwon et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib29)and retain the original audio without enhancement\. This preserves the original acoustic characteristics and avoids potential alterations to language\-specific phonetic cues that may be crucial for underrepresented languages\.

#### Speaker Diarization and Segmentation\.

We then apply thepyannote\-Precision\-2[Bredin \(2017\)](https://arxiv.org/html/2610.19388#bib.bib18);[Bredin et al\. \(2020\)](https://arxiv.org/html/2610.19388#bib.bib19);[Bredin \(2023\)](https://arxiv.org/html/2610.19388#bib.bib20)speaker diarization pipeline to partition each recording into speaker\-homogeneous regions\. Overlaps shorter than 0\.5 seconds are retained, as they often reflect natural conversational backchannels rather than speaker interference\. For segments exceeding 30 seconds, we applySilero VAD[Team \(2024\)](https://arxiv.org/html/2610.19388#bib.bib16)and recursively split them at the longest silence until the duration constraint is satisfied\. Detailed segmentation algorithms are provided in Appendix[2](https://arxiv.org/html/2610.19388#alg2)\.

### 3\.4Normalization and Filtering

For both subsets, we apply the following common normalization and filtering procedures to ensure consistency across the corpus\.

#### Speech Normalization\.

First, all the speech segments are resampled to 24 kHz and encoded as 16\-bit PCM FLAC format\. We then apply RMS\-based loudness normalization to−23\-23dBFS, with the gain constrained to\[−6,\+6\]\[\-6,\+6\]dB to prevent excessive amplification or attenuation, followed by a peak ceiling of−1\-1dBFS to prevent clipping\.

#### Metadata Normalization\.

Next, we standardize language metadata across ELAR collections, where naming conventions vary widely\. Language labels are assigned based on collection\-level metadata and subsequently normalized using the Ethnologue333https://www\.ethnologue\.com/language catalog, mapping aliases and naming variants to canonical language identifiers\.

#### Data Filtering\.

Finally, we remove sign language recordings, which typically contain little or no spoken content, and discard samples with no detectable speech usingSilero VAD\.

## 4Experiments

### 4\.1Long\-Tail Speech Modeling

Table 3:Monolingual and multilingual ASR performance across 30 pivot languages of ELARSpeech\. Languages are sorted by data duration within each category, and we report CER \(%\) for every result\. Best and second\-best results are shown in bold and underlined, respectively\. Macro CER is averaged across languages, while win rate denotes the proportion of languages for which each model achieves the lowest CER\.#### Experimental Setup\.

We first investigate whether ELARSpeech can extend the language coverage of existing speech models into the linguistic long tail\. Both ASR and TTS are evaluated in monolingual and multilingual settings\. Monolingual adaptation focuses on 30 pivot languages across resource levels for model comparisons \(Appendix[F](https://arxiv.org/html/2610.19388#A6)\), whereas the multilingual counterpart jointly trains a single model on all 233 languages represented across the train, development, and test splits\. Models are initially trained on human\-annotated ELARSpeech\-L, with language\-specific Whisper models additionally fine\-tuned for all 233 languages to pseudo\-label ELARSpeech\-U\. Pseudo\-labeled scaling is then examined by progressively augmenting the training data with pseudo\-labeled ELARSpeech\-U\.

#### Models\.

For ASR, we fine\-tune two distinct modeling paradigms: Whisper\-large\-v3[Radford et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib1), an encoder\-decoder model, and OmniASR\-CTC\-v2[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4), a CTC\-based[Graves et al\. \(2006\)](https://arxiv.org/html/2610.19388#bib.bib21)model\. For TTS, we fine\-tune OmniVoice[Zhu et al\. \(2026\)](https://arxiv.org/html/2610.19388#bib.bib5), a diffusion LLM\-based[Yang et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib24);[Gong et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib25)multilingual model\. Following prior findings on the importance of data quality in low\-resource TTS[Ogun et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib23), we restrict TTS training data to utterances with both SNR and C50 above 20 dB\.

#### Evaluation Metrics\.

For ASR, we report character error rate \(CER\) against human transcriptions\. For TTS, we evaluate intelligibility using CER from monolingual Whisper models and perceptual quality using mean opinion score \(MOS\), following the evaluation protocol of[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3)\. For evaluation across all 233 languages, we report UTMOS[Baba et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib30)as an automatic proxy\. Details of the MOS evaluation protocol are provided in Appendix[G](https://arxiv.org/html/2610.19388#A7)\.

#### Training Details\.

All models are optimized using AdamW[Loshchilov and Hutter \(2017\)](https://arxiv.org/html/2610.19388#bib.bib22)until convergence, and the checkpoint achieving the best development\-set performance is selected for final evaluation\. Detailed training configurations and hyperparameter settings are provided in Appendix[H](https://arxiv.org/html/2610.19388#A8)\.

### 4\.2Long\-Tail Representation Analysis

#### Evaluation Protocol\.

We subsequently investigate how well existing multilingual S3Ms generalize to the linguistic long tail represented in ELARSpeech\. Following the ML\-SUPERB protocol[Shi et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib43), we evaluate these models under 10\-minute and 1\-hour supervision\. We consider four tasks from the original benchmark: monolingual ASR, multilingual ASR, language identification \(LID\), and joint ASR with LID\.

#### Models\.

We benchmark five multilingual S3Ms: XLS\-R[Babu et al\. \(2021\)](https://arxiv.org/html/2610.19388#bib.bib46), w2v\-BERT 2\.0[Barrault et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib45), MMS[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3), XEUS[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13), and OmniASR\-W2V[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4), together with mel\-filterbank features as a baseline\.

#### Evaluation Metrics\.

Consistent with the original ML\-SUPERB protocol, we report CER for ASR and accuracy for LID, with all aggregate results reported as macro averages across languages\.

#### Training Details\.

We follow the original ML\-SUPERB training recipe for all experiments\.

## 5Results

### 5\.1Automatic Speech Recognition

#### Monolingual Adaptation\.

Table[3](https://arxiv.org/html/2610.19388#S4.T3)demonstrates that ELARSpeech\-L enables ASR adaptation across diverse long\-tail languages\. In the monolingual setting, Whisper achieves the lowest CER for 18 of the 30 pivot languages, with a macro CER of 22\.48\. OmniASR\-CTC is more competitive on languages without ISO 639\-3 codes, as the 300M model attains a macro CER of 24\.50, which is comparable to Whisper’s 24\.63 despite having substantially fewer parameters\. However, Whisper remains ahead on ISO\-coded languages \(20\.33 vs\. 24\.18\)\. This trend is consistent with prior findings on the effectiveness of CTC\-based models in low\-resource settings[Kim et al\. \(2017\)](https://arxiv.org/html/2610.19388#bib.bib34);[Moritz et al\. \(2019\)](https://arxiv.org/html/2610.19388#bib.bib35);[Nayeem et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib36), suggesting that their advantages may extend further into the linguistic long tail\. For OmniASR\-CTC variants, scaling model size from 300M to 1B parameters improves performance on 16 of the 30 languages, but the gains are not consistent across languages\. Specifically, all four languages with over 60 hours of training data benefit from the larger model, while results remain mixed for lower\-resource languages\.

#### Multilingual Knowledge Transfer\.

In contrast, massively multilingual adaptation shows signs of linguistic interference, as the multilingual model increases macro CER from 22\.48 to 30\.71, outperforming the monolingual counterpart on only one of the 30 pivot languages\. This suggests that scaling the number of languages alone does not necessarily yield positive cross\-lingual transfer, when supervision is limited and distributed across heterogeneous long\-tail languages, such as diverse language families and recording environments\.

002525505075751001003333343435353636Pseudo\-labeled data retained \(%\)CER \(%\)↓\\downarrowFigure 2:ASR performance with increasing amounts of pseudo\-labeled ELARSpeech\-U\.
#### Pseudo\-Labeled Scaling\.

Figure[2](https://arxiv.org/html/2610.19388#S5.F2)shows how pseudo\-labeled ELARSpeech\-U affects ASR performance across 233 languages\. Specifically, we compute the confidence of each pseudo\-labeled utterance as the mean confidence across its predicted tokens and apply language\-specific quartile cutoffs to vary the amount of retained data\. Across these settings, augmenting ELARSpeech\-L with pseudo\-labeled speech consistently improves transcription performance, achieving up to a 6\.1% relative reduction in CER\. However, retaining the entire pseudo\-labeled corpus does not yield the best performance, with moderate confidence filtering leading to greater improvements\. This suggests that effective scaling with unlabeled speech depends not only on quantity but also on pseudo\-label quality\.

Table 4:Cross\-domain ASR performance on ELARSpeech and external test sets: CV26 and INEL\.
#### Cross\-Domain Robustness\.

To assess whether the benefits of ELARSpeech extend beyond its source archive, we conduct a cross\-domain ASR evaluation on four languages from two external corpora[Ardila et al\. \(2020\)](https://arxiv.org/html/2610.19388#bib.bib6);[Baranova \(2025\)](https://arxiv.org/html/2610.19388#bib.bib38)\. We compare models trained on the original external data, augmented with ELARSpeech\-L, and further augmented with pseudo\-labeled ELARSpeech\-U, evaluating each on both domains\. As shown in Table[4](https://arxiv.org/html/2610.19388#S5.T4), external\-only models perform notably worse on ELARSpeech, indicating a considerable domain mismatch between ELARSpeech and existing speech corpora\. Moreover, augmenting the training data with ELARSpeech generally improves performance on the external domains, with gains observed in three of the four languages\. Together, these results suggest that ELARSpeech broadens domain coverage without compromising performance on the original domains\.

### 5\.2Text\-to\-Speech

#### Low\-Resource Adaptation\.

LanguageCodeData\(H\)ReferenceMonolingualMultilingual\(Ground Truth\)OmniVoiceOmniVoiceCER↓\\downarrowMOS↑\\uparrowCER↓\\downarrowMOS↑\\uparrowCER↓\\downarrowMOS↑\\uparrowLanguages with ISO 639\-3 CodesC\. Maselamxz124\.928\.394\.0414\.713\.8812\.873\.79Y\. Mixtecxty64\.88\.704\.124\.753\.584\.833\.82Tzeltaltzh19\.823\.203\.7612\.583\.3312\.933\.45Iñapariinp13\.111\.373\.9415\.433\.6014\.973\.59Araonaaro7\.723\.433\.7515\.523\.0416\.653\.26E\. Minyagemq6\.019\.843\.8825\.783\.5427\.663\.96Ju\|’hoansiktz5\.142\.933\.4841\.602\.9046\.392\.94Z\. Chatinoczn5\.110\.734\.065\.453\.465\.943\.55Macro Average21\.073\.8816\.983\.4217\.783\.55Win Rate––75\.0%25\.0%25\.0%75\.0%Languages without ISO 639\-3 CodesJitotoltec–12\.613\.313\.969\.173\.5210\.203\.79C\. Zoque–10\.117\.114\.1311\.243\.5911\.793\.82Caijia–6\.428\.354\.4026\.983\.9627\.724\.17S\. L\. P\. C\.–6\.016\.894\.026\.623\.887\.003\.69Zok–5\.238\.353\.6938\.483\.2738\.713\.42Muyu–5\.023\.583\.8814\.803\.1915\.843\.06Macro Average22\.934\.0117\.883\.5718\.543\.66Win Rate––100\.0%33\.3%0\.0%66\.7%

Table 5:Monolingual and multilingual TTS performance on pivot languages with at least 5 hours of training data, with MOS ratings from non\-native listeners\. Multi\-word language names are abbreviated\.Table[5](https://arxiv.org/html/2610.19388#S5.T5)presents TTS results across the 14 pivot languages with at least 5 hours of training data after SNR/C50 filtering\. Despite the limited amount of transcribed speech available for many languages, the models yield intelligible speech across a diverse set of low\-resource languages, including those with only a few hours of data\. In several cases, synthesized speech even achieves lower CER than the ground\-truth recordings, providing further evidence of strong intelligibility under the ASR\-based evaluation, although this may partly reflect the more standardized acoustic characteristics of synthesized speech relative to real\-world field recordings\. Overall, these results demonstrate that ELARSpeech can support TTS adaptation even for languages with extremely limited speech resources\. Compared with the monolingual setting, the multilingual counterpart yields slightly higher CER overall but improves MOS for the majority of evaluated languages, suggesting a trade\-off between intelligibility and perceptual quality\.

#### Pseudo\-Labeled Scaling\.

1818202022222424CER↓\\downarrow

002525505075751001002\.552\.552\.652\.652\.752\.75Pseudo\-labeled data retained \(%\)UTMOS↑\\uparrow

Figure 3:TTS performance with increasing amounts of pseudo\-labeled ELARSpeech\-U\.ModelData \(H\)\# Langs\.\# ELAR\# Params\.Mono\. ASRMulti\. ASRMulti\. LIDMulti\. ASR \+ LIDScoreCER↓\\downarrowCER↓\\downarrowAcc\.↑\\uparrowAcc\.↑\\uparrowCER↓\\downarroww2v\-BERT 2\.0†4\.5M143–600M41\.0 / 29\.224\.6 / 20\.370\.0 / 86\.883\.2 / 90\.624\.2 / 20\.3–XLS\-R†436K128–300M39\.7 / 30\.629\.3 / 22\.066\.9 / 87\.955\.6 / 85\.628\.4 / 22\.9–1B40\.5 / 30\.930\.4 / 26\.370\.9 / 85\.866\.4 / 87\.128\.6 / 25\.2–MMS†491K1,406–300M33\.8 / 30\.528\.7 / 24\.062\.3 / 84\.371\.9 / 74\.331\.5 / 30\.0–1B33\.3 / 25\.721\.3 / 18\.184\.8 / 86\.173\.3 / 74\.826\.0 / 25\.5–XEUS†1M4,057–600M30\.3 / 25\.121\.1 / 20\.181\.5 / 87\.386\.4 / 91\.322\.9 / 19\.6–Mel\-Filterbank––––71\.6 / 64\.973\.8 / 70\.26\.4 / 8\.530\.9 / 43\.373\.0 / 69\.60\.0 / 0\.0w2v\-BERT 2\.04\.5M1436 \(4\.2%\)600M53\.8 / 45\.857\.6/51\.750\.6 /57\.754\.1 / 79\.457\.7 / 51\.9901\.9/ 944\.2XLS\-R436K1287 \(5\.5%\)300M53\.5 / 45\.759\.2 / 53\.149\.3 / 56\.549\.4 / 81\.358\.9 / 52\.8848\.5 / 921\.71B52\.1/44\.958\.9 / 53\.049\.9 / 55\.951\.9 /82\.259\.2 / 52\.2880\.8 / 935\.2MMS491K1,406333 \(23\.7%\)300M53\.6 / 46\.058\.4 / 52\.753\.7/61\.954\.2 / 81\.058\.6 / 52\.3901\.6 / 949\.81B51\.9/44\.656\.9/51\.554\.4/ 57\.755\.3/83\.156\.7/50\.8968\.4/980\.1XEUS1M4,057342 \(8\.4%\)600M52\.2 / 45\.060\.2 / 54\.449\.4 / 57\.663\.6/ 82\.159\.8 / 53\.5898\.4 / 915\.6OmniASR\-W2V4\.3M1,690140 \(8\.3%\)300M52\.1/ 45\.558\.9 / 52\.745\.5 / 41\.949\.5 / 79\.858\.5 / 52\.4854\.3 / 857\.81B52\.3 / 45\.157\.9 / 51\.949\.0 /59\.352\.1 / 81\.057\.7/51\.3901\.4 /967\.8

Table 6:ML\-SUPERB\-style evaluation on ELARSpeech\. Results are reported under 10\-minute / 1\-hour supervision\. Data and \# Langs\. denote the amount and reported language coverage of SSL pretraining data\. \# ELAR denotes the number of languages in each model’s publicly available language inventory that overlap with ELARSpeech, with the percentage relative to its reported language coverage shown in parentheses\. Gray italicized rows \(†\\dagger\) show the original \(non\-long\-tail\) ML\-SUPERB results reported in[Chen et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib13)for reference\.Figure[3](https://arxiv.org/html/2610.19388#S5.F3)shows the effect of scaling pseudo\-labeled ELARSpeech\-U on TTS performance\. Unlike ASR, increasing the amount of pseudo\-labeled speech does not improve intelligibility, with CER instead showing a slight increase as more pseudo\-labeled data are added\. In contrast, UTMOS initially improves with additional pseudo\-labeled speech, reaches its highest score at an intermediate retention level, and declines when the entire pseudo\-labeled corpus is retained\. This suggests that pseudo\-labeled speech can provide useful additional acoustic variation for TTS, improving perceptual quality at moderate scale despite a slight degradation in intelligibility, while lower\-confidence pseudo\-labels diminish these gains as more data are retained\.

### 5\.3Representation Analysis

#### Low\-Resource Representation Transfer\.

As shown in Table[6](https://arxiv.org/html/2610.19388#S5.T6), all pretrained representations consistently outperform the mel\-filterbank baseline across the evaluated tasks, indicating that multilingual pretraining remains beneficial for the long\-tail languages in ELARSpeech\. These improvements persist under both supervision budgets, suggesting that pretrained representations can provide useful features even when language\-specific supervision is extremely limited\.

#### Beyond Existing Benchmarks\.

However, despite these consistent gains over the baseline, the relative advantages among pretrained representations become less pronounced on ELARSpeech\. For example, 10\-minute monolingual ASR CER ranges from 30\.3 to 41\.0 on ML\-SUPERB \(26\.1% relative difference\), compared with only 51\.9 to 53\.8 \(3\.5% relative difference\) on ELARSpeech\. Meanwhile, all models perform consistently worse on ELARSpeech than on ML\-SUPERB under matched supervision budgets\. These results show that existing S3Ms exhibit weaker absolute performance on ELARSpeech, losing much of the relative advantage observed on the original benchmark across high\-resource languages\.

#### Pretraining Scale, Coverage, and Exposure\.

Interestingly, these differences are not well explained by pretraining scale or nominal language coverage alone\. While XEUS has the largest absolute overlap with ELARSpeech, this represents only a small fraction \(8\.4%\) of its broader language inventory; in contrast, MMS exhibits a considerably higher proportion \(23\.7%\) of ELARSpeech languages and achieves the strongest overall performance\. This suggests that effective long\-tail transfer may depend less on simply scaling the number of pretraining languages and more on maintaining balanced exposure to low\-resource languages, highlighting the need for linguistic long\-tail resources to drive further advances in multilingual speech models\.

## 6Conclusion

In this work, we introduced ELARSpeech, a large\-scale speech corpus comprising 30,202 hours across 675 underrepresented languages\. Beyond its scale, ELARSpeech extends multilingual speech resources further into the linguistic long tail while preserving the acoustic and domain diversity of real\-world field recordings\. Subsequently, our experiments demonstrate that ELARSpeech supports effective ASR and TTS adaptation across diverse underrepresented languages, with pseudo\-labeled data providing further opportunities for scaling beyond limited human annotations on long\-tail languages\. At the same time, representation analysis shows that pretraining scale and language coverage alone are insufficient predictors of long\-tail transfer, highlighting the importance of speech resources that enable balanced exposure to low\-resource languages\. We anticipate ELARSpeech to serve as a foundation for advancing multilingual speech technology further into the linguistic long tail\.

## Limitations

As ELARSpeech is constructed from archival collections, its language and recording metadata may contain inconsistencies or uncertainties inherited from the original deposits\. While our processing pipeline standardizes and filters the collected data, it cannot fully resolve inaccuracies in archival metadata\. Additionally, our MOS evaluation relies primarily on non\-native listeners due to the difficulty of recruiting native speakers across hundreds of low\-resource languages\. Accordingly, the evaluation focuses on general perceptual qualities such as naturalness, artifacts, and listening comfort rather than linguistic accuracy or native\-like pronunciation and prosody\. We complement these human judgments with automatic measures of intelligibility and perceptual quality, but this aspect remains an important direction for future work\.

## Ethical Statement

ELARSpeech is derived from archival recordings of endangered and underrepresented language communities, for which responsible use requires particular attention to the rights and conditions associated with the source materials\. Prior to data collection, we consulted the ELAR team regarding the use of archived materials for machine\-learning research and follow their guidance on access, attribution, and downstream use\. We include only materials designated asOpenorUser Access, excludeSubscriber Accessmaterials that require permission from collection stewards, and preserve collection\-level provenance and attribution information\. We recognize that permission to access archival materials does not remove the responsibility to respect the individuals and communities represented in them\. We acknowledge the agency of language communities in determining how their linguistic resources are used and commit to honoring requests to remove materials from ELARSpeech\. Accordingly, ELARSpeech is intended for non\-commercial research, with downstream use subject to the applicable licensing and collection\-specific conditions\.

## References

- Ardilaet al\.\(2020\)R\. Ardila, M\. Branson, K\. Davis, M\. Kohler, J\. Meyer, M\. Henretty, R\. Morais, L\. Saunders, F\. Tyers, and G\. WeberCommon Voice: a massively\-multilingual speech corpus\.InProceedings of the Language Resources and Evaluation Conference,pp\. 4218–4222\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p3.1),[§2\.1](https://arxiv.org/html/2610.19388#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2610.19388#S5.SS1.SSS0.Px4.p1.1)\.
- Asonitiset al\.\(2026\)A\. Asonitis, L\. A\. Lanzendörfer, F\. Berdoz, and R\. WattenhoferWorldSpeech: a multilingual speech corpus from around the world\.arXiv preprint arXiv:2605\.09167\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p4.1),[§2\.1](https://arxiv.org/html/2610.19388#S2.SS1.p1.1)\.
- Babaet al\.\(2024\)K\. Baba, W\. Nakata, Y\. Saito, and H\. SaruwatariThe t05 system for the VoiceMOS Challenge 2024: transfer learning from deep image classifier to naturalness MOS prediction of high\-quality synthetic speech\.InIEEE Spoken Language Technology Workshop \(SLT\),pp\. 818–824\.External Links:[Document](https://dx.doi.org/10.1109/SLT61566.2024.10832315)Cited by:[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px3.p1.1)\.
- Babuet al\.\(2021\)A\. Babu, C\. Wang, A\. Tjandra, K\. Lakhotia, Q\. Xu, N\. Goyal, K\. Singh, P\. Von Platen, Y\. Saraf, J\. Pino,et al\.XLS\-r: self\-supervised cross\-lingual speech representation learning at scale\.arXiv preprint arXiv:2111\.09296\.Cited by:[§4\.2](https://arxiv.org/html/2610.19388#S4.SS2.SSS0.Px2.p1.1)\.
- Baranova \(2025\)V\. BaranovaINEL kalmyk corpus\.External Links:[Document](https://dx.doi.org/10.25592/uhhfdm.17676),[Link](https://doi.org/10.25592/uhhfdm.17676)Cited by:[§5\.1](https://arxiv.org/html/2610.19388#S5.SS1.SSS0.Px4.p1.1)\.
- Barraultet al\.\(2023\)L\. Barrault, Y\. Chung, M\. C\. Meglioli, D\. Dale, N\. Dong, M\. Duppenthaler, P\. Duquenne, B\. Ellis, H\. Elsahar, J\. Haaheim,et al\.Seamless: multilingual expressive and streaming speech translation\.arXiv preprint arXiv:2312\.05187\.Cited by:[§4\.2](https://arxiv.org/html/2610.19388#S4.SS2.SSS0.Px2.p1.1)\.
- Boulianne \(2022\)G\. BouliannePhoneme transcription of endangered languages: an evaluation of recent asr architectures in the single speaker scenario\.InFindings of the Association for Computational Linguistics: ACL 2022,pp\. 2301–2308\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p5.1)\.
- Bredinet al\.\(2020\)H\. Bredin, R\. Yin, J\. M\. Coria, G\. Gelly, P\. Korshunov, M\. Lavechin, D\. Fustes, H\. Titeux, W\. Bouaziz, and M\. GillPyannote\. audio: neural building blocks for speaker diarization\.InICASSP 2020\-2020 IEEE International conference on acoustics, speech and signal processing \(ICASSP\),pp\. 7124–7128\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px3.p1.1)\.
- Bredin \(2017\)H\. BredinPyannote\. metrics: a toolkit for reproducible evaluation, diagnostic, and error analysis of speaker diarization systems\.hypothesis100\(60\),pp\. 90\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px3.p1.1)\.
- Bredin \(2023\)H\. BredinPyannote\. audio 2\.1 speaker diarization pipeline: principle, benchmark, and recipe\.In24th Interspeech Conference \(INTERSPEECH 2023\),pp\. 1983–1987\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px3.p1.1)\.
- Chenet al\.\(2022\)S\. Chen, C\. Wang, Z\. Chen, Y\. Wu, S\. Liu, Z\. Chen, J\. Li, N\. Kanda, T\. Yoshioka, X\. Xiao,et al\.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:[Appendix I](https://arxiv.org/html/2610.19388#A9.p1.1)\.
- Chenet al\.\(2024\)W\. Chen, W\. Zhang, Y\. Peng, X\. Li, J\. Tian, J\. Shi, X\. Chang, S\. Maiti, K\. Livescu, and S\. WatanabeTowards robust speech representation learning for thousands of languages\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 10205–10224\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p2.1),[§1](https://arxiv.org/html/2610.19388#S1.p4.1),[§1](https://arxiv.org/html/2610.19388#S1.p5.1),[§2\.2](https://arxiv.org/html/2610.19388#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2610.19388#S3.SS1.SSS0.Px3.p1.1),[§4\.2](https://arxiv.org/html/2610.19388#S4.SS2.SSS0.Px2.p1.1),[Table 6](https://arxiv.org/html/2610.19388#S5.T6)\.
- Conneauet al\.\(2023\)A\. Conneau, M\. Ma, S\. Khanuja, Y\. Zhang, V\. Axelrod, S\. Dalmia, J\. Riesa, C\. Rivera, and A\. BapnaFLEURS: few\-shot learning evaluation of universal representations of speech\.InIEEE Spoken Language Technology Workshop,pp\. 798–805\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p3.1),[§2\.1](https://arxiv.org/html/2610.19388#S2.SS1.p1.1)\.
- Desplanqueset al\.\(2020\)B\. Desplanques, J\. Thienpondt, and K\. DemuynckEcapa\-tdnn: emphasized channel attention, propagation and aggregation in tdnn based speaker verification\.arXiv preprint arXiv:2005\.07143\.Cited by:[Appendix I](https://arxiv.org/html/2610.19388#A9.p1.1)\.
- Gonget al\.\(2025\)S\. Gong, S\. Agarwal, Y\. Zhang, J\. Ye, L\. Zheng, M\. Li, C\. An, P\. Zhao, W\. Bi, J\. Han,et al\.Scaling diffusion language models via adaptation from autoregressive models\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 5046–5073\.Cited by:[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1)\.
- Graveset al\.\(2006\)A\. Graves, S\. Fernández, F\. Gomez, and J\. SchmidhuberConnectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks\.InProceedings of the 23rd international conference on Machine learning,pp\. 369–376\.Cited by:[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1)\.
- Guzmánet al\.\(2026\)D\. Guzmán, L\. H\. Beyene, J\. O\. Alabi, Y\. Jeon, D\. Klakow, and D\. I\. AdelaniOpenBibleTTS: large\-scale speech resources and tts models for low\-resource languages\.arXiv preprint arXiv:2606\.09553\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p4.1),[§2\.1](https://arxiv.org/html/2610.19388#S2.SS1.p1.1)\.
- Heet al\.\(2024\)H\. He, Z\. Shang, C\. Wang, X\. Li, Y\. Gu, H\. Hua, L\. Liu, C\. Yang, J\. Li, P\. Shi,et al\.Emilia: an extensive, multilingual, and diverse speech dataset for large\-scale speech generation\.In2024 IEEE Spoken Language Technology Workshop \(SLT\),pp\. 885–890\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px2.p1.1)\.
- Kerenet al\.\(2025\)G\. Keren, A\. Kozhevnikov, Y\. Meng, C\. Ropers, M\. Setzler, S\. Wang, I\. Adebara, M\. Auli, C\. Balioglu,et al\.Omnilingual asr: open\-source multilingual speech recognition for 1600\+ languages\.arXiv preprint arXiv:2511\.09690\.Cited by:[Appendix H](https://arxiv.org/html/2610.19388#A8.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2610.19388#S1.p1.1),[§1](https://arxiv.org/html/2610.19388#S1.p3.1),[§1](https://arxiv.org/html/2610.19388#S1.p5.1),[§2\.1](https://arxiv.org/html/2610.19388#S2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2610.19388#S3.SS1.SSS0.Px3.p1.1),[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1),[§4\.2](https://arxiv.org/html/2610.19388#S4.SS2.SSS0.Px2.p1.1)\.
- Kimet al\.\(2017\)S\. Kim, T\. Hori, and S\. WatanabeJoint ctc\-attention based end\-to\-end speech recognition using multi\-task learning\.In2017 IEEE international conference on acoustics, speech and signal processing \(ICASSP\),pp\. 4835–4839\.Cited by:[§5\.1](https://arxiv.org/html/2610.19388#S5.SS1.SSS0.Px1.p1.1)\.
- Kwonet al\.\(2026\)K\. Kwon, J\. So, and S\. LeeCoreaSpeech: korean speech corpus via jamo\-based coreset selection for efficient and robust korean speech generation\.Advances in Neural Information Processing Systems38\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px2.p1.1)\.
- Lauet al\.\(2025\)M\. Lau, Q\. Chen, Y\. Fang, T\. Xu, T\. Chen, and P\. GolikData quality issues in multilingual speech datasets: the need for sociolinguistic awareness and proactive language planning\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 7466–7492\.Cited by:[Appendix B](https://arxiv.org/html/2610.19388#A2.p1.1)\.
- Lavechinet al\.\(2023\)M\. Lavechin, M\. Métais, H\. Titeux, A\. Boissonnet, J\. Copet, M\. Rivière, E\. Bergelson, A\. Cristia, E\. Dupoux, and H\. BredinBrouhaha: multi\-task training for voice activity detection, speech\-to\-noise ratio, and c50 room acoustics estimation\.In2023 IEEE Automatic Speech Recognition and Understanding Workshop \(ASRU\),pp\. 1–7\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px2.p1.1)\.
- Le Ferrandet al\.\(2025\)É\. Le Ferrand, C\. M\. B\. Hauser, J\. Hartshorne, and E\. Prud’hommeauxFaithful transcription: leveraging bible recordings to improve asr for endangered languages\.InProceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia\-Pacific Chapter of the Association for Computational Linguistics,pp\. 333–342\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p5.1)\.
- Leet al\.\(2023\)M\. Le, A\. Vyas, B\. Shi, B\. Karrer, L\. Sari, R\. Moritz, M\. Williamson, V\. Manohar, Y\. Adi, J\. Mahadeokar,et al\.Voicebox: text\-guided multilingual universal speech generation at scale\.Advances in neural information processing systems36,pp\. 14005–14034\.Cited by:[Appendix I](https://arxiv.org/html/2610.19388#A9.p1.1)\.
- Liet al\.\(2023\)X\. Li, S\. Takamichi, T\. Saeki, W\. Chen, S\. Shiota, and S\. WatanabeYodas: youtube\-oriented dataset for audio and speech\.In2023 IEEE Automatic Speech Recognition and Understanding Workshop \(ASRU\),pp\. 1–8\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p3.1)\.
- Li and Niehues \(2026\)Z\. Li and J\. NiehuesMultimodal in\-context learning for asr of low\-resource languages\.InFindings of the Association for Computational Linguistics: ACL 2026,pp\. 24745–24760\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p5.1)\.
- Loshchilov and Hutter \(2017\)I\. Loshchilov and F\. HutterDecoupled weight decay regularization\.arXiv preprint arXiv:1711\.05101\.Cited by:[Appendix H](https://arxiv.org/html/2610.19388#A8.SS0.SSS0.Px3.p1.1),[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px4.p1.1)\.
- Maet al\.\(2024\)L\. Ma, D\. Guo, K\. Song, Y\. Jiang, S\. Wang, L\. Xue, W\. Xu, H\. Zhao, B\. Zhang, and L\. XieWenetspeech4tts: a 12,800\-hour mandarin tts corpus for large speech generation model benchmark\.arXiv preprint arXiv:2406\.05763\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px2.p1.1)\.
- Meyeret al\.\(2022\)J\. Meyer, D\. I\. Adelani, E\. Casanova, A\. Öktem, D\. W\. J\. Weber, S\. Kabongo, E\. Salesky, I\. Orife, C\. Leong, P\. Ogayo,et al\.Bibletts: a large, high\-fidelity, multilingual, and uniquely african speech corpus\.arXiv preprint arXiv:2207\.03546\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p4.1)\.
- Moritzet al\.\(2019\)N\. Moritz, T\. Hori, and J\. Le RouxStreaming end\-to\-end speech recognition with joint ctc\-attention based models\.In2019 IEEE Automatic Speech Recognition and Understanding Workshop \(ASRU\),pp\. 936–943\.Cited by:[§5\.1](https://arxiv.org/html/2610.19388#S5.SS1.SSS0.Px1.p1.1)\.
- Nayeemet al\.\(2025\)M\. Nayeem, M\. S\. Tabrej, K\. J\. Deb, S\. Goswami, and M\. A\. HakimAutomatic speech recognition in the modern era: architectures, training, and evaluation\.arXiv preprint arXiv:2510\.12827\.Cited by:[§5\.1](https://arxiv.org/html/2610.19388#S5.SS1.SSS0.Px1.p1.1)\.
- Ogunet al\.\(2023\)S\. Ogun, V\. Colotte, and E\. VincentCan we use common voice to train a multi\-speaker tts system?\.In2022 IEEE Spoken Language Technology Workshop \(SLT\),pp\. 900–905\.Cited by:[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1)\.
- Ogunet al\.\(2024\)S\. Ogun, A\. T\. Owodunni, T\. Olatunji, E\. Alese, B\. Oladimeji, T\. Afonja, K\. Olaleye, N\. A\. Etori, and T\. Adewumi1000 african voices: advancing inclusive multi\-speaker multi\-accent speech synthesis\.arXiv preprint arXiv:2406\.11727\.Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px2.p1.1)\.
- Pfistereret al\.\(2026\)S\. Pfisterer, F\. Grötschla, L\. Lanzendörfer, F\. Yan, and R\. WattenhoferEurospeech: a multilingual speech corpus\.Advances in Neural Information Processing Systems38\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p4.1)\.
- Pratapet al\.\(2024\)V\. Pratap, A\. Tjandra, B\. Shi, P\. Tomasello, A\. Babu, S\. Kundu, A\. Elkahky, Z\. Ni, A\. Vyas, M\. Fazel\-Zarandi,et al\.Scaling speech technology to 1,000\+ languages\.Journal of Machine Learning Research25\(97\),pp\. 1–52\.Cited by:[Appendix H](https://arxiv.org/html/2610.19388#A8.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2610.19388#S1.p1.1),[§1](https://arxiv.org/html/2610.19388#S1.p2.1),[§1](https://arxiv.org/html/2610.19388#S1.p3.1),[§1](https://arxiv.org/html/2610.19388#S1.p5.1),[§2\.2](https://arxiv.org/html/2610.19388#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2610.19388#S3.SS1.SSS0.Px3.p1.1),[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px3.p1.1),[§4\.2](https://arxiv.org/html/2610.19388#S4.SS2.SSS0.Px2.p1.1)\.
- Radfordet al\.\(2023\)A\. Radford, J\. W\. Kim, T\. Xu, G\. Brockman, C\. McLeavey, and I\. SutskeverRobust speech recognition via large\-scale weak supervision\.InInternational conference on machine learning,pp\. 28492–28518\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p1.1),[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1)\.
- Shiet al\.\(2023\)J\. Shi, D\. Berrebbi, W\. Chen, H\. Chung, E\. Hu, W\. P\. Huang, X\. Chang, S\. Li, A\. Mohamed, H\. Lee,et al\.Ml\-superb: multilingual speech universal performance benchmark\.arXiv preprint arXiv:2305\.10615\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p5.1),[§4\.2](https://arxiv.org/html/2610.19388#S4.SS2.SSS0.Px1.p1.1)\.
- Shiet al\.\(2024\)J\. Shi, S\. Wang, W\. Chen, M\. Bartelds, V\. B\. Kumar, J\. Tian, X\. Chang, D\. Jurafsky, K\. Livescu, H\. Lee,et al\.Ml\-superb 2\.0: benchmarking multilingual speech models across modeling constraints, languages, and datasets\.arXiv preprint arXiv:2406\.08641\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p5.1)\.
- Snyder \(1997\)J\. P\. SnyderFlattening the earth: two thousand years of map projections\.University of Chicago press\.Cited by:[1\(a\)](https://arxiv.org/html/2610.19388#S2.F1.sf1)\.
- Team \(2024\)S\. TeamSilero vad: pre\-trained enterprise\-grade voice activity detector \(vad\), number detector and language classifier\.GitHub\.Note:[https://github\.com/snakers4/silero\-vad](https://github.com/snakers4/silero-vad)Cited by:[§3\.3](https://arxiv.org/html/2610.19388#S3.SS3.SSS0.Px3.p1.1)\.
- Valk and Alumäe \(2021\)J\. Valk and T\. AlumäeVoxLingua107: a dataset for spoken language recognition\.In2021 IEEE spoken language technology workshop \(SLT\),pp\. 652–658\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p3.1)\.
- Wanget al\.\(2021\)C\. Wang, M\. Riviere, A\. Lee, A\. Wu, C\. Talnikar, D\. Haziza, M\. Williamson, J\. Pino, and E\. DupouxVoxPopuli: a large\-scale multilingual speech corpus for representation learning, semi\-supervised learning and interpretation\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 993–1003\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p4.1)\.
- Yanget al\.\(2025\)A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1)\.
- Zhanget al\.\(2023\)Y\. Zhang, W\. Han, J\. Qin, Y\. Wang, A\. Bapna, Z\. Chen, N\. Chen, B\. Li, V\. Axelrod, G\. Wang,et al\.Google usm: scaling automatic speech recognition beyond 100 languages\.arXiv preprint arXiv:2303\.01037\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p1.1)\.
- Zhuet al\.\(2026\)H\. Zhu, L\. Ye, W\. Kang, Z\. Yao, L\. Guo, F\. Kuang, Z\. Han, W\. Zhuang, L\. Lin, and D\. PoveyOmnivoice: towards omnilingual zero\-shot text\-to\-speech with diffusion language models\.arXiv preprint arXiv:2604\.00688\.Cited by:[§1](https://arxiv.org/html/2610.19388#S1.p1.1),[§4\.1](https://arxiv.org/html/2610.19388#S4.SS1.SSS0.Px2.p1.1)\.

## Appendix

## Appendix ALicense and Terms of Use

As mentioned in Section[3\.1](https://arxiv.org/html/2610.19388#S3.SS1), ELARSpeech is derived from materials archived in the Endangered Languages Archive \(ELAR\)\. Prior to data collection and processing, we contacted the ELAR team to clarify the permitted use of archived materials for machine\-learning research\. ELAR confirmed that such use, including model training and research outputs, is permissible provided that downstream use complies with the applicable licensing and access conditions of the source materials\.

We restrict ELARSpeech to materials designated asOpenorUser Access, excluding allSubscriber Accessmaterials, which involve additional permission procedures and may require explicit authorization from individual collection stewards beyond access approval from ELAR\. For each included collection, we preserve its provenance, including the ELAR permanent identifier and collection steward information where available, as well as the applicable access and licensing conditions\.

ELARSpeech consists of machine\-learning\-oriented derivatives produced through our preprocessing pipeline and is released under the Creative Commons Attribution\-NonCommercial\-ShareAlike 4\.0 International \(CC BY\-NC\-SA 4\.0\) license\. This license does not supersede the ownership, copyright, attribution requirements, or access conditions of the original ELAR materials\. Downstream use and redistribution must comply with both the ELARSpeech license and the applicable collection\-specific terms, while preserving the attribution and provenance of the source collections\.

Figure 4:Language family distribution from 78 unique language families of ELARSpeech\.
## Appendix BDataset Design Principles

The source archival collections were created for language documentation, where language identities, varieties, and orthographic conventions are determined by the source collections and their communities\. Given that language labels, orthographic conventions, and language\-dialect boundaries can introduce ambiguity in massively multilingual speech datasets, particularly for low\-resource languages[Lau et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib37), we follow several principles to avoid imposing additional assumptions during dataset construction\.

#### Preservation Over Normalization\.

First, we preserve the language names, variety distinctions, and orthographic conventions provided by the original collections whenever possible\. In particular, we do not normalize different orthographies or merge language varieties solely to obtain a more uniform representation\. This approach prioritizes preserving the linguistic distinctions in the source archive over enforcing consistency across collections\.

#### No Forced ISO Mapping\.

Second, we retain languages and varieties without ISO 639\-3 codes rather than assigning them to potentially related ISO\-coded languages\. This approach avoids imposing standardized identifiers that may not accurately reflect the linguistic distinctions in the original collections\. Consequently, the reported language coverage includes both ISO\-coded entries and those identified only by their original language names\.

#### Minimal Linguistic Intervention\.

Third, we do not redefine whether two documented varieties constitute distinct languages or dialects, nor do we systematically merge entries based on linguistic similarity\. Such distinctions can depend on sociolinguistic, historical, and community\-specific factors that cannot be reliably inferred from the recordings alone\. We therefore defer to the distinctions represented in the source collections rather than introducing a new linguistic taxonomy during data processing\.

#### Known Limitations\.

Aforementioned principles preserve the structure of the original archive but do not eliminate ambiguities already present in archival metadata\. Language names may vary across collections, boundaries between languages and dialects may be sociolinguistically complex, and orthography may differ within or across collections\. Accordingly, the language counts reported for ELARSpeech should be interpreted as reflecting the documented language and variety identities represented in the source archive, rather than as an independently verified linguistic taxonomy\.

## Appendix CDetailed Statistics of ELARSpeech

Algorithm 1Automatic Transcription Tier Selection and Heuristic Scoring Rules\(a\) Transcription tier selection

1:ELAN file set

ℰ\\mathcal\{E\}, acceptance threshold

τa=70\\tau\_\{a\}=70, review threshold

τr=50\\tau\_\{r\}=50
2:Accepted tiers

𝒯a\\mathcal\{T\}\_\{a\}and tiers requiring manual review

𝒯r\\mathcal\{T\}\_\{r\}
3:

𝒯a←∅\\mathcal\{T\}\_\{a\}\\leftarrow\\emptyset
4:

𝒯r←∅\\mathcal\{T\}\_\{r\}\\leftarrow\\emptyset
5:for all

e∈ℰe\\in\\mathcal\{E\}do

6:

𝒞e←∅\\mathcal\{C\}\_\{e\}\\leftarrow\\emptyset
7:for alltier

t∈et\\in edo

8:

nt←Normalize​\(TierID​\(t\)\)n\_\{t\}\\leftarrow\\textsc\{Normalize\}\(\\textsc\{TierID\}\(t\)\)
9:

lt←Normalize​\(LinguisticType​\(t\)\)l\_\{t\}\\leftarrow\\textsc\{Normalize\}\(\\textsc\{LinguisticType\}\(t\)\)
10:

pt←Normalize​\(ParentRef​\(t\)\)p\_\{t\}\\leftarrow\\textsc\{Normalize\}\(\\textsc\{ParentRef\}\(t\)\)
11:

s⁡\(t\)←ScoreTier​\(nt,lt,pt\)s\(t\)\\leftarrow\\textsc\{ScoreTier\}\(n\_\{t\},l\_\{t\},p\_\{t\}\)
12:

s⁡\(t\)←min⁡\(100,max⁡\(0,s⁡\(t\)\)\)s\(t\)\\leftarrow\\min\(100,\\max\(0,s\(t\)\)\)
13:

𝒞e←𝒞e∪\{\(t,s⁡\(t\)\)\}\\mathcal\{C\}\_\{e\}\\leftarrow\\mathcal\{C\}\_\{e\}\\cup\\\{\(t,s\(t\)\)\\\}
14:endfor

15:

t∗←arg⁡maxt∈𝒞e⁡s⁡\(t\)t^\{\\ast\}\\leftarrow\\arg\\max\_\{t\\in\\mathcal\{C\}\_\{e\}\}s\(t\)
16:if

s⁡\(t∗\)≥τas\(t^\{\\ast\}\)\\geq\\tau\_\{a\}then

17:

𝒯a←𝒯a∪\{\(e,t∗\)\}\\mathcal\{T\}\_\{a\}\\leftarrow\\mathcal\{T\}\_\{a\}\\cup\\\{\(e,t^\{\\ast\}\)\\\}
18:elseif

s⁡\(t∗\)≥τrs\(t^\{\\ast\}\)\\geq\\tau\_\{r\}then

19:

𝒯r←𝒯r∪\{\(e,t∗\)\}\\mathcal\{T\}\_\{r\}\\leftarrow\\mathcal\{T\}\_\{r\}\\cup\\\{\(e,t^\{\\ast\}\)\\\}
20:else

21:RejectCandidate\(

e,t∗e,t^\{\\ast\}\)

22:endif

23:endfor

\(b\) Heuristic scoring rules

Scoring\.All applicable rules are accumulated, and the resulting score is clipped to\[0,100\]\[0,100\]\. For labels such astx@Speaker1, the substring preceding@is additionally compared with conventional source\-text labels\.

We provide additional statistics on the linguistic, resource, and acoustic characteristics of ELARSpeech\. Specifically, Figure[4](https://arxiv.org/html/2610.19388#A1.F4)presents the number of languages and total speech hours across 78 language families, illustrating the broad linguistic diversity and resource distribution of the corpus\. Additionally, Table[7](https://arxiv.org/html/2610.19388#A9.T7)provides detailed statistics for 606 canonical languages out of 675 languages, excluding the multilingual subset whose samples cannot be unambiguously assigned to a single language\. The table reports the total duration per split for each language\.

## Appendix DAutomatic Annotation Tier Selection

As described in Section[3\.2](https://arxiv.org/html/2610.19388#S3.SS2), ELAR’s annotation files exhibit heterogeneous tier naming and contain multiple annotation types, including transcriptions, translations, glosses, and metadata, requiring automatic identification of valid transcription tiers\.

To automatically identify time\-aligned transcription tiers, we assign each annotation tier a heuristic score based on its tier ID, linguistic type, and parent\-tier information\. The scoring function incorporates both positive indicators associated with source transcriptions \(e\.g\.,transcription,tx, andutterance\) and negative indicators associated with auxiliary annotations \(e\.g\.,translation,gloss, andnote\)\. Scores from all applicable rules are accumulated and clipped to the range\[0,100\]\[0,100\]\. For each ELAN file, the highest\-scoring tier is selected as the candidate transcription tier\.

Algorithm[1](https://arxiv.org/html/2610.19388#alg1)summarizes the complete tier selection procedure and heuristic scoring rules\. Candidate tiers with scores of at least 70 are accepted automatically, while those scoring between 50 and 70 are manually verified\. Candidates scoring below 50 are discarded\. This combination of heuristic ranking and selective manual verification enables us to accommodate heterogeneous annotation conventions across ELAR collections without requiring manual inspection of every annotation file\.

Specifically, across 39,338 unique annotation tiers, 7,144 received scores of 70 or higher, 3,936 received scores between 50 and 70, and 28,258 received scores below 50\. We automatically accept tiers scoring at least 70, manually review those scoring between 50 and 70, and reject those scoring below 50\. Manual validation estimated a 99\.3% precision for automatically accepted tiers, with 993 of 1,000 randomly sampled tiers confirmed as valid transcriptions, while 99\.9% of automatically rejected tiers were confirmed to be invalid \(999 of 1,000\)\. Among the 3,936 intermediate\-score tiers, manual review identified 1,035 as valid transcription tiers \(26\.3%\)\.

## Appendix EUnlabeled Data Segmentation

Algorithm 2Clip Extraction with Recursive Split1:Merged segments

ℳ\\mathcal\{M\}, audio root

𝒟in\\mathcal\{D\}\_\{\\mathrm\{in\}\}, output root

𝒟out\\mathcal\{D\}\_\{\\mathrm\{out\}\},

τmin=0\.3\\tau\_\{\\min\}=0\.3s,

T=30T=30s, VAD threshold

θ\\theta
2:Final clips

ℱ\\mathcal\{F\}, excluded set

𝒳\\mathcal\{X\}
3:

ℱ←∅\\mathcal\{F\}\\leftarrow\\emptyset,

𝒳←∅\\mathcal\{X\}\\leftarrow\\emptyset
4:// Stage 1: discard short spans, write initial clips

5:for allrecording

ffwith

𝒢f⊆ℳ\\mathcal\{G\}\_\{f\}\\subseteq\\mathcal\{M\}do

6:

𝒜←\{s∈𝒢f:d⁡\(s\)\>τmin\}\\mathcal\{A\}\\leftarrow\\\{s\\in\\mathcal\{G\}\_\{f\}:d\(s\)\>\\tau\_\{\\min\}\\\}
7:

𝒳←𝒳∪\(𝒢f∖𝒜\)\\mathcal\{X\}\\leftarrow\\mathcal\{X\}\\cup\(\\mathcal\{G\}\_\{f\}\\setminus\\mathcal\{A\}\),

ℒ←∅\\mathcal\{L\}\\leftarrow\\emptyset
8:for all

s∈𝒜s\\in\\mathcal\{A\}do

9:WriteClip\(

f,s,𝒟outf,s,\\mathcal\{D\}\_\{\\mathrm\{out\}\}\)

10:if

d⁡\(s\)≤Td\(s\)\\leq Tthen

ℱ←ℱ∪\{s\}\\mathcal\{F\}\\leftarrow\\mathcal\{F\}\\cup\\\{s\\\}else

ℒ←ℒ∪\{s\}\\mathcal\{L\}\\leftarrow\\mathcal\{L\}\\cup\\\{s\\\}
11:endfor

12:// Stage 2: recursively split clips longer thanTT

13:if

ℒ≠∅\\mathcal\{L\}\\neq\\emptysetthen

𝒱←SileroVAD​\(f,θ\)\\mathcal\{V\}\\leftarrow\\textsc\{SileroVAD\}\(f,\\theta\)
14:for all

s∈ℒs\\in\\mathcal\{L\}do

15:

𝒫←SplitSegment​\(s,𝒱,T,τmin\)\\mathcal\{P\}\\leftarrow\\textsc\{SplitSegment\}\(s,\\mathcal\{V\},T,\\tau\_\{\\min\}\)
16:WriteClip\(

f,p,𝒟outf,p,\\mathcal\{D\}\_\{\\mathrm\{out\}\}\) for all

p∈𝒫p\\in\\mathcal\{P\}
17:

ℱ←ℱ∪𝒫\\mathcal\{F\}\\leftarrow\\mathcal\{F\}\\cup\\mathcal\{P\}
18:endfor

19:endfor

20:functionSplitSegment\(

s,𝒱,T,τmins,\\mathcal\{V\},T,\\tau\_\{\\min\}\)

21:if

d⁡\(s\)≤Td\(s\)\\leq Tthenreturn

\{s\}\\\{s\\\}
22:

c←c\\leftarrowlongest silence in

𝒱∩s\\mathcal\{V\}\\cap swith

≥τmin\\geq\\tau\_\{\\min\}on both sides; else nearest VAD boundary; else midpoint of

ss
23:returnSplitSegment\(

s<cs\_\{<c\}\)

∪\\cupSplitSegment\(

s\>cs\_\{\>c\}\)

24:endfunction

Algorithm[2](https://arxiv.org/html/2610.19388#alg2)describes the procedure used to transform long\-form recordings into training\-ready clips for ELARSpeech\-U\. Following the diarization post\-processing described in Section[3\.3](https://arxiv.org/html/2610.19388#S3.SS3), the resulting speaker\-homogeneous segments are used as input to the segmentation pipeline\.

We first discard segments shorter than or equal to 0\.3 seconds, as such short clips are unlikely to contain sufficient speech content for reliable model training\. Subsequently, segments of at most 30 seconds are directly retained, while longer segments are recursively split usingSilero VAD\.

For each segment exceeding 30 seconds, we prioritize the longest silence that leaves at least 0\.3 seconds of audio on both sides\. If no silence exists, we use the nearest valid VAD boundary; if no valid boundary is available, the segment midpoint is used as a fallback\. This procedure is recursively applied until all resulting clips are at most 30 seconds long\. Overlap information is retained as metadata to enable flexible filtering in downstream use\.

## Appendix FPivot Language Selection

We select 30 pivot languages to provide a representative evaluation across diverse resource levels and linguistic groups\.

First, we divide ELARSpeech languages into three resource bins based on the amount of available training data: high\-resource \(\>20\>20hours\), mid\-resource \(10\-20 hours\), and low\-resource \(<10<10hours\), and sample languages across these bins to ensure diversity in resource availability\.

Second, given our primary focus on long\-tail and previously underrepresented languages, we select 15 of the 30 pivot languages from languages without an ISO 639\-3 code\. We also balance these languages across resource bins to avoid concentrating the evaluation on a particular resource level\.

Finally, we maximize language\-family diversity within the resulting language pool\. When multiple selected languages belong to the same language family, we replace overlapping languages, where possible, with candidates from different families with comparable amounts of training data\. This procedure yields 30 pivot languages spanning diverse resource levels and language families, while maintaining balanced representation of languages that are rarely covered by existing benchmarks\.

![Refer to caption](https://arxiv.org/html/2610.19388v1/appendix/mos.png)Figure 5:Instruction page provided to participants\.
## Appendix GMOS Evaluation Protocol

#### Participants\.

We recruited 36 listeners through community outreach\. As the TTS evaluation covers 14 extremely low\-resource languages that participants were not expected to speak, listeners were instructed to assess onlyacoustic naturalness and listening quality, considering human\-likeness, audible artifacts \(e\.g\., noise, glitches, metallic timbre, muffling, or dropouts\), and overall listening comfort\. They were explicitly instructed not to consider intelligibility, pronunciation or intonation, or familiarity with the language, which were evaluated separately using automatic metrics\. Ratings were collected on a 5\-point scale\. The detailed evaluation instructions are shown in Figure[5](https://arxiv.org/html/2610.19388#A6.F5)\.

#### Evaluation Procedure\.

We constructed a pool of 210 utterances by sampling five clips for each of 42 system\-language pairs \(three systems×\\times14 languages\)\. Each utterance was rated by 9 listeners\. Participants were required to listen in a quiet environment over headphones or earphones; loudspeakers were not allowed, to reduce variability in acoustic presentation\.

## Appendix HTraining Details

#### Automatic Speech Recognition\.

For monolingual ASR, Whisper\-large\-v3 and OmniASR\-CTC\-v2 are fine\-tuned for 25 and 50 epochs, respectively, with a batch size of 32\. For multilingual ASR, Whisper\-large\-v3 is fine\-tuned for 150K steps with a batch size of 128\. The two models also require different strategies to accommodate languages unsupported during pretraining\. Whisper uses a BPE\-based vocabulary together with language\-ID tokens\. Since its BPE vocabulary can represent text in the newly introduced languages, no vocabulary expansion is required\. For languages without a pretrained language\-ID token, we introduce a new language token and warm\-start its embedding from that of the<en\>language token before fine\-tuning on the target language\. In contrast, OmniASR\-CTC\-v2 uses a fixed output vocabulary for CTC decoding\. When the training data contain tokens absent from its original vocabulary, we extend the vocabulary with the missing tokens and expand the corresponding linear head, randomly initializing the newly added rows\.

#### Text\-to\-Speech\.

Unlike the utterance\-based batching used for ASR, OmniVoice employs packed\-sequence training to improve training efficiency\. Accordingly, monolingual TTS models are fine\-tuned for 50 epochs with a total batch duration of 2\.5 minutes, while the multilingual OmniVoice model is fine\-tuned for 100K steps\. Similar to ASR, we adapt OmniVoice to languages unsupported during pretraining by introducing a new language token for each language and initializing its embedding from the<en\>token, as its BPE vocabulary can represent text in the newly introduced languages without expansion\.

#### Optimization\.

All models are optimized using the AdamW[Loshchilov and Hutter \(2017\)](https://arxiv.org/html/2610.19388#bib.bib22)optimizer, with a tri\-stage learning rate schedule on 8 H100 GPUs\. The warmup, hold, and decay phases follow a ratio of 0\.1, 0\.5, and 0\.4, respectively\. The learning rate increases from5​e−75e\{\-7\}to1​e−51e\{\-5\}during warmup, remains at1​e−51e\{\-5\}during the hold phase, and decays to5​e−75e\{\-7\}\. For multilingual training, we adopt the language sampling strategy of[Pratap et al\. \(2024\)](https://arxiv.org/html/2610.19388#bib.bib3);[Keren et al\. \(2025\)](https://arxiv.org/html/2610.19388#bib.bib4), with sampling probabilities defined aspl∝\(nl/N\)βLp\_\{l\}\\propto\(n\_\{l\}/N\)^\{\\beta\_\{L\}\}\. Here,nln\_\{l\}denotes the number of samples for languagell, andNNis the total number of samples\. We setβL=0\.5\\beta\_\{L\}=0\.5\.

## Appendix IFull Evaluation Results

Tables[8](https://arxiv.org/html/2610.19388#A9.T8)and[9](https://arxiv.org/html/2610.19388#A9.T9)present the full ASR and TTS evaluation results, respectively, across the 233 languages included in all three data splits \(train, development, and test\)\. For the full TTS evaluation, we additionally report speaker similarity \(SIM\-O\), computed as the cosine similarity between speaker embeddings of synthesized speech and ground\-truth utterances extracted using the WavLM\-based ECAPA\-TDNN model[Desplanques et al\. \(2020\)](https://arxiv.org/html/2610.19388#bib.bib32);[Chen et al\. \(2022\)](https://arxiv.org/html/2610.19388#bib.bib31), following VoiceBox[Le et al\. \(2023\)](https://arxiv.org/html/2610.19388#bib.bib33)\.

LanguageISO 639\-3Labeled \(H\)Unlabeled \(H\)LanguageISO 639\-3Labeled \(H\)Unlabeled \(H\)TrainDevTestTrainDevTestAbuiabz10\.970\.641\.32129\.08Ainuain0\.000\.000\.002\.78Ainu \(Saru\)aib0\.000\.000\.006\.28Aiomeaki0\.000\.000\.000\.42Aisimmq0\.000\.000\.005\.16Ajumbumuc0\.270\.000\.0062\.98Akanaka0\.490\.000\.001\.44Akhaahk5\.490\.330\.642\.29Akpesibe10\.800\.631\.25107\.75Aleutale0\.000\.000\.00263\.10Alta, Northernaqn2\.830\.200\.35118\.97Amami\-Oshima, Southernams0\.700\.000\.00100\.80Amba \(Suwa\)rwm0\.000\.000\.0064\.12Ambelwgo5\.530\.290\.62137\.85Amurdakamg0\.000\.000\.0048\.70Analanm8\.520\.491\.016\.99Angaitéaqt1\.630\.100\.1815\.07Animereanf4\.980\.260\.5477\.32Anindilyakwaaoi0\.000\.000\.009\.97Antia Whistling Language–0\.000\.000\.000\.83Anusauq0\.000\.000\.006\.82Apalaíapy0\.000\.000\.005\.88Apurinãapu3\.300\.180\.449\.92Arabic, Judeo\-Iraqiyhd2\.350\.140\.289\.68Arandic \(Family\)–0\.000\.000\.002\.43Araonaaro11\.380\.711\.40414\.79Arapahoarp14\.400\.841\.7251\.04Arotei6\.770\.370\.7826\.87Artaatz0\.040\.000\.0015\.95Ashéninka, Perenéprq0\.000\.000\.0018\.16Assyrian Neo\-Aramaicaii0\.710\.000\.0082\.59Asuriasr0\.000\.000\.0011\.68Avatimeavn3\.460\.210\.4267\.33Awiakay–3\.160\.180\.4118\.07Awutuafu1\.830\.100\.2229\.96Axambahb0\.000\.000\.00112\.27Ayoreoayo1\.870\.080\.21134\.72Ayutla Mixtecmiy0\.000\.000\.0081\.45Babankibbk1\.590\.000\.002\.31Babar, Northbcd0\.030\.000\.0030\.24Babar, Southeastvbb0\.100\.000\.0031\.90Badagabfq0\.000\.000\.001\.66Bafiaksf0\.000\.000\.0037\.63Bafutbfd0\.090\.000\.000\.48Bai, Panyibfc0\.030\.000\.0030\.69Bajjikavjk0\.000\.000\.001\.57Balochibal0\.000\.000\.0022\.39Banam Bayvrt9\.130\.531\.1029\.56Bandialbqj9\.240\.561\.1111\.37Bandial \(Kujireray\)bqj1\.170\.000\.000\.09Baniwabwi7\.840\.510\.8726\.08Barambrd18\.881\.142\.232\.27Bareinbva1\.740\.090\.2015\.78Baskeet–0\.230\.000\.0023\.72Bastimentos Creole English–8\.580\.510\.990\.91Batharibhm0\.000\.000\.00132\.96Batibtc9\.280\.511\.08240\.67Baunibpe0\.000\.000\.0024\.00Bayotbda0\.000\.000\.005\.32Baïnounk Gubëëher–0\.150\.000\.002\.70Beezenbnz0\.000\.000\.003\.98Belaitbeg14\.640\.901\.6542\.07Betta Kurumbaxub0\.000\.000\.005\.53Bihibh2\.100\.110\.262\.29Bikyabyb0\.000\.000\.004\.59Binebon2\.900\.170\.3264\.73Biturmcc0\.000\.000\.000\.52Blablangablp6\.430\.370\.7894\.85Bogoŋ–0\.000\.000\.0066\.47Bolivian Guaraní–0\.000\.000\.005\.97Bombmf0\.010\.000\.0014\.59Bom\-Kimbmf0\.240\.000\.0013\.05Bongubpu1\.420\.100\.169\.95Brahuibrh1\.110\.000\.0012\.58Brembuq1\.590\.110\.171\.09Bribribzd0\.000\.000\.0045\.91Bullom Sobuy0\.160\.000\.001\.40Bumbmv5\.470\.300\.61118\.09Bunakbfn5\.600\.330\.670\.53Burarrabvr4\.420\.280\.5569\.17Burarra \(Gun\-nartpa\)bvr6\.130\.360\.6487\.66Busuubju0\.000\.000\.0011\.80Buulfb0\.690\.000\.0021\.61Caacmsq0\.960\.000\.000\.84Cabiyarícbb1\.580\.120\.1912\.34Cabécarcjp0\.000\.000\.0012\.90Cahuillachl2\.440\.170\.315\.06Caijia–10\.010\.551\.17160\.65Calderón Highland Quichuaqud19\.111\.152\.2412\.78Camsákbh4\.620\.280\.5222\.34Caquintecot0\.000\.000\.00102\.62Central \!Kung \(Mangetti Dune\)–0\.710\.000\.000\.00Chamalalcji0\.010\.000\.0036\.65Chatino, Tataltepeccta0\.480\.000\.005\.10Chatino, Zacatepecctz0\.030\.000\.00277\.18Chechenche0\.000\.000\.002\.22Cheyi–0\.680\.000\.0018\.29Chiapas Zoque–12\.690\.781\.467\.51Chimalapa Zoquezoh9\.800\.621\.130\.24Chimborazo Highland Quichuaqug14\.600\.881\.787\.34Chiniafi5\.740\.340\.6839\.27Chiquitano \(Bésiro\)cax1\.620\.110\.198\.84Chorote \(Family\)–0\.000\.000\.008\.40Chukchickt0\.000\.000\.0026\.19Chulym \(Melet\)clw1\.060\.000\.00120\.14Chulym \(Tutal\)clw0\.000\.000\.0010\.76Chácobocao20\.921\.252\.51237\.61Cicipuawc6\.000\.340\.69146\.24Cinenicie1\.660\.000\.0064\.69Cofáncon5\.680\.310\.7221\.38Cora \(Family\)–4\.100\.240\.5172\.10Cubeocub0\.440\.000\.003\.39Curripacokpc2\.600\.170\.295\.70Cuvokcuv1\.040\.000\.0013\.57Dahalodal0\.000\.000\.007\.44Dalabonngk0\.000\.000\.0063\.27Damakawadam0\.000\.000\.000\.96Danaudnu0\.000\.000\.007\.45Dao–0\.000\.000\.0026\.67Dargwa \(Chirag\)dar35\.292\.104\.0880\.95Dariprs0\.000\.000\.0043\.77Datoogatcc2\.060\.120\.240\.00Datooga \(Asimjeeg\)tcc27\.351\.583\.2795\.09Dawoodidmk7\.590\.430\.9079\.96Desanodes0\.000\.000\.008\.04Dghwededgh1\.740\.000\.0049\.42Dhaonfa0\.000\.000\.0012\.26Djauandjn0\.000\.000\.004\.10Doedoe0\.000\.000\.000\.57Dolgandlg17\.451\.021\.970\.00Domarirmt0\.000\.000\.0036\.72Drungduu0\.220\.000\.0034\.15Dwangnnu0\.790\.000\.002\.66Dzajen0\.000\.000\.00164\.56Dzardzongke–5\.350\.320\.6930\.80Dzongkhadzo0\.000\.000\.0015\.26Dâwkwa7\.170\.370\.8430\.90Eastern Huasteca Nahuatlnhe0\.000\.000\.003\.69Ehueunehu0\.060\.000\.0023\.85Eibelaail8\.900\.511\.04393\.93Emaemmw5\.510\.300\.6634\.79Emplawasemw37\.452\.254\.51265\.91Enawené\-Nawéunk3\.940\.250\.4522\.51Engaenq0\.000\.000\.0076\.30Englisheng0\.000\.000\.005\.19Enxetenx2\.260\.130\.2727\.07Ersuers1\.750\.100\.20306\.47Ese Ejjaese0\.000\.000\.0023\.43Etonetn0\.000\.000\.0024\.68Eveneve12\.840\.781\.560\.53E’ñapa Woromaipupbh13\.350\.791\.5341\.08Fanbakfnb0\.080\.000\.0025\.65Fangfak0\.020\.000\.0011\.50Faniafni3\.230\.180\.3690\.53Farefaregur0\.000\.000\.0035\.11Fasfqs0\.710\.000\.002\.79Fonfon0\.000\.000\.004\.13Fungwaula4\.290\.260\.5040\.73Gadaba, Bodogbj9\.470\.571\.1576\.60Gadjerawanggdh0\.220\.000\.0096\.25Gandalug0\.450\.000\.007\.22Gantsgao0\.000\.000\.000\.50Gargarnaye–6\.030\.360\.680\.26Garifunacab2\.680\.190\.3318\.48Gavião of Jiparanágvo0\.000\.000\.0047\.50Gbanzirigbg0\.000\.000\.007\.10Germandeu0\.000\.000\.000\.32Ghanonggaghn5\.590\.330\.6714\.43Ghulfanghl0\.440\.000\.0011\.20Gitxsangit1\.330\.000\.003\.00Glavdaglw0\.000\.000\.009\.54Glio\-Oubioub0\.000\.000\.001\.45Goemaiank1\.320\.000\.000\.00Golpalja0\.840\.000\.0042\.34Gongdukgoe2\.620\.150\.2936\.28Gorapgoq0\.090\.000\.005\.93Gorwaagow73\.294\.288\.67641\.39Great Andamanese, Mixedgac0\.000\.000\.0072\.00Griko–0\.000\.000\.001\.61Guarayugyr3\.620\.180\.4313\.94Guatógta4\.140\.240\.4628\.87Guduf\-Gavagdf0\.170\.000\.000\.00Guernésiaisnrf2\.300\.130\.275\.25Gulf Arabicafb5\.960\.360\.7513\.48Gumut,ească–1\.110\.000\.001\.29Gunguw0\.000\.000\.001\.54Gurindji Kriolgjr0\.350\.000\.001\.40Gwenogwe7\.820\.490\.9561\.46Gyalsumdogyo0\.000\.000\.006\.87Gyeligyi0\.000\.000\.0010\.71Hadzahts4\.230\.240\.49424\.80Hahonhah0\.000\.000\.0019\.43Hahon \(Ratsua\)hah0\.000\.000\.003\.92Hakka Chinesehak0\.000\.000\.0094\.57Harsusihss0\.790\.000\.0031\.93Heiyohoauk4\.910\.250\.5752\.63Hereroher0\.000\.000\.004\.31Highland Popolucapoi5\.750\.330\.643\.97Hixkaryánahix0\.000\.000\.0017\.79Hmong Dawmww0\.100\.000\.001\.77Hobyóthoh0\.000\.000\.0052\.51Horpaero0\.450\.000\.0038\.67Hrusohru2\.380\.150\.28130\.41Huastec \(South Eastern\)hus0\.480\.000\.0025\.12Huillichehuh8\.200\.510\.9668\.92Hupahup0\.020\.000\.000\.54Ibatanivb16\.220\.951\.8474\.31Ibibioibb1\.580\.100\.214\.83Igboibo0\.000\.000\.000\.98Imbabura Highland Quichuaqvi13\.890\.831\.646\.58Imroingimr0\.460\.000\.0019\.86Inari Saamismn0\.000\.000\.0027\.88Ingainb1\.590\.000\.0013\.59Ingrianizh2\.950\.160\.33459\.54Ipikoipo0\.000\.000\.00121\.75Ipiliipi0\.000\.000\.002\.05Irántxe \(Myky\)irn0\.190\.000\.0046\.37Isanzuisn49\.182\.925\.80351\.44Itelmenitl0\.790\.000\.000\.15Iteneite0\.000\.000\.000\.84Ixcatecixc1\.030\.000\.0060\.45Ixilixl0\.000\.000\.0014\.29Iñapariinp25\.551\.503\.05122\.83Jabutíjbt6\.510\.370\.75101\.01Jawejaz7\.820\.420\.900\.67Jejueojje3\.600\.210\.45179\.81Jitotoltec–18\.561\.082\.103\.96Jowshaqani–2\.050\.110\.2223\.60Ju\|’hoansiktz79\.904\.719\.48267\.56Júmajua1\.610\.090\.2019\.25Kaaporurb3\.950\.200\.45172\.38Kabardiankbd1\.160\.000\.002\.79Kadazan Dusundtp1\.370\.000\.0055\.83Kagulukki0\.000\.000\.000\.11Kaikekzq2\.990\.190\.3515\.02Kainji \(Family\)–0\.000\.000\.0052\.96Kakabekke4\.150\.260\.5224\.96Kakataibo\-Kashibocbr3\.410\.190\.40123\.56Kalmyk\-Oiratxal2\.430\.140\.2828\.41Kalmyk\-Oirat \(Durvud\)xal1\.940\.110\.2124\.59Kamkdx16\.020\.961\.8677\.91Kamasxas34\.281\.994\.000\.00Kanamaríknm0\.300\.000\.0025\.91Kandozi\-Chapracbu2\.680\.150\.3221\.05Kangjiakxs1\.750\.110\.2266\.53Kanum, Smärkykxq0\.570\.000\.000\.00Karagaskim5\.350\.320\.6471\.17Karara–0\.000\.000\.0027\.47Karbimjw1\.390\.000\.002\.19Kateviribsh0\.000\.000\.004\.02Katla \(Julud\)kcr0\.000\.000\.002\.44Kelabitkzi11\.470\.701\.351\.78Kerinci \(Koto Keras\)–3\.220\.200\.365\.34Kgalagadixkv17\.281\.002\.1541\.11Khamyangksu0\.000\.000\.0018\.55Khanty \(Eastern\)kca0\.000\.000\.0060\.81Khengkhaxkf0\.000\.000\.0032\.19Khoekhoenaq0\.000\.000\.000\.77Khoibu Nagankb10\.500\.621\.2354\.25Kholosi–0\.000\.000\.001\.70Kibala \(Lubolo\)blv0\.000\.000\.0065\.89Kichwa de Chibuleo–2\.180\.130\.26197\.76Kimbukiv0\.240\.000\.0024\.20Kinnauri, Chitkulicik3\.050\.180\.354\.27Kiongkkm0\.190\.000\.000\.99Klonkyo0\.000\.000\.0012\.33Kolekl4\.420\.240\.5116\.69Komoxom1\.260\.000\.0049\.13Kondaknd0\.120\.000\.0046\.27Konomalakoa15\.110\.871\.763\.92Koranakqz0\.000\.000\.001\.08Koreguajecoe0\.000\.000\.0023\.78Korokxr0\.000\.000\.00166\.15Koryakkpy24\.951\.472\.89312\.24Koshinkid0\.640\.000\.0025\.51Kovekvc0\.800\.000\.009\.24Koyeekkt0\.000\.000\.001\.67Kpashampbn8\.080\.480\.9659\.78Kukeght2\.930\.180\.3611\.82Kulatpg4\.030\.240\.4827\.26Kumarbhag Pahariakmj0\.000\.000\.0022\.45Kumiaidih0\.930\.000\.000\.42Kuna, San Blascuk0\.000\.000\.0031\.45Kunbarlangwlg1\.540\.000\.0067\.92Kundal Shahishd2\.840\.140\.310\.53Kungkfl0\.000\.000\.006\.56Kung\-Ekokaknw0\.000\.000\.003\.21Kursavfaj0\.500\.000\.005\.16Kurtokhaxkz0\.000\.000\.0057\.55Kurudukjr0\.340\.000\.00216\.91Kushikuh0\.000\.000\.0018\.15Kusua–0\.000\.000\.00185\.43Kwakwb4\.390\.260\.51115\.44Kwak’walakwk0\.100\.000\.00121\.48Kwatl–0\.450\.000\.000\.00Kweguxwg0\.000\.000\.000\.71K’iche’quc15\.300\.941\.774\.84Lacandonlac0\.000\.000\.0015\.57Lachilbt0\.000\.000\.0026\.71Laitu Chinclj0\.000\.000\.0027\.21Lakurumaulxm12\.650\.821\.466\.82Lavatbura\-Lamusonglbv2\.710\.180\.3135\.16Lazlzz2\.220\.120\.28205\.38Lelepalpa3\.950\.250\.4710\.49Letilti22\.221\.292\.693\.56Lezgilez0\.570\.000\.000\.77Lhokpulhp0\.000\.000\.0080\.19Likilio0\.400\.000\.0058\.51Limassabme1\.150\.000\.0092\.52Longgulgu0\.000\.000\.002\.57Lundayehlnd1\.850\.000\.002\.74Macaguajemcl0\.000\.000\.008\.43Macowpc0\.140\.000\.008\.22Maemme1\.820\.080\.180\.00Mafeamkv11\.570\.601\.260\.83Magimgu0\.000\.000\.004\.89Mah Merimhe0\.000\.000\.003\.38Makurápmpu7\.280\.440\.86121\.40Mal Pahariamkb1\.580\.000\.0060\.47Malaccan Portuguese Creolemcm0\.280\.000\.003\.24Malengpkt0\.680\.000\.0050\.10Maléku Jaíkagut0\.000\.000\.00218\.11Manatpmr0\.400\.000\.007\.99Mandate0\.780\.000\.0014\.47Manepmkr0\.960\.000\.002\.43Mansimns0\.000\.000\.000\.57Maquiritarimch1\.700\.000\.0013\.31Maragusmrs4\.460\.270\.5150\.22Maring Naganng0\.000\.000\.008\.11Masela, Centralmxz136\.797\.9516\.0335\.61Masela, Eastvme0\.010\.000\.0030\.49Masela, Westmss0\.020\.000\.0030\.20Massepmvs22\.471\.362\.68373\.42Matlatzinca, San Franciscomat19\.561\.142\.2948\.25Matsesmcf0\.000\.000\.0062\.25Matukar Panaumjk16\.530\.931\.92144\.11Maungmph0\.000\.000\.0028\.91Mazandaranimzn0\.000\.000\.001\.56Mbukushumhw0\.000\.000\.000\.82Mbyá Guaranígun1\.740\.130\.1815\.82Meakambut–0\.000\.000\.009\.47Mehrigdq5\.110\.280\.6360\.46Meiteimni0\.000\.000\.001\.59Mele\-Filamxe9\.460\.541\.175\.61Mewahang, Easternemg0\.530\.000\.0031\.72Mewahang, Westernraf17\.180\.961\.96115\.08Mezquital Otomiote3\.740\.220\.4378\.82Miahuatlán Zapoteczam0\.000\.000\.0019\.25Minyag, Easternemq12\.490\.731\.4268\.05Minyag, Westernwmg8\.060\.460\.9917\.17Misakgum0\.000\.000\.0039\.73Mixtec \(Family\)–0\.000\.000\.001\.94Mochomhc9\.590\.561\.17113\.57Mocovímoc1\.060\.000\.0070\.77Mongolian, Peripheral \(Jirim\)mvf8\.780\.501\.0097\.42Monpa, Tawangtwm0\.000\.000\.0056\.29Monsang Naganmh0\.610\.000\.006\.27Mormhz2\.570\.170\.285\.72Mororimok2\.380\.140\.293\.09Mulaogiu18\.641\.102\.133\.31Mundabli\-Mufuboe0\.000\.000\.0091\.65Mungbammij1\.090\.000\.00194\.32Mungbam \(Ngun\)mij2\.180\.130\.252\.74Mursimuz0\.000\.000\.000\.63Murut, Bookanbnb14\.580\.851\.8033\.88Muyu–6\.300\.380\.7426\.26Myenemye3\.440\.210\.4122\.03Méniktnr0\.220\.000\.004\.19Nakimff0\.490\.000\.00106\.91Nalunaj0\.000\.000\.0035\.33Nalögonlz2\.300\.150\.2668\.07Namuyinmy0\.000\.000\.008\.75Nar Phunpa0\.000\.000\.002\.23Nar Phu \(Nar\)npa0\.030\.000\.008\.64Nasapbb0\.290\.000\.0027\.78Nasalnsy0\.320\.000\.00281\.90Nawurinaw0\.000\.000\.003\.03Nda’nda’nnz0\.730\.000\.0053\.38Negidalneg0\.400\.000\.0033\.77Nenets \(Tundra\)yrk0\.000\.000\.000\.45Nese–2\.180\.110\.264\.87Ngaanyatjarrantj0\.000\.000\.004\.56Ngan’gityemerrinam1\.080\.000\.009\.08Nheengatuyrl4\.100\.250\.5011\.44Nigeria Mambilamzk1\.560\.000\.0038\.51Nihalinll1\.060\.000\.0018\.57Nilanil0\.130\.000\.0017\.89Nindemwi0\.600\.000\.0031\.59Nivaclécag0\.840\.000\.0023\.29Nkaminkq8\.370\.561\.0415\.55Nkoroonkx0\.130\.000\.0017\.36Noonsnf1\.580\.000\.001\.60North Ambrymmmg1\.410\.000\.0048\.23Northern Altai \(Teleut\)atv2\.950\.200\.3565\.28Notrebly4\.200\.240\.5256\.56Nsiteubko7\.890\.510\.83285\.94Nuernus0\.000\.000\.003\.96Nyenkhaneh0\.500\.000\.001\.69N\|\|ng \(Western\)ngh0\.000\.000\.0025\.08N’kepsku0\.070\.000\.0027\.36Oinam–0\.000\.000\.00115\.92Okiekoki15\.330\.911\.802\.66Olekhaole0\.000\.000\.0026\.93Olekha \(Northeastern\)ole0\.000\.000\.0077\.27Ongotabxe0\.050\.000\.002\.80Opolgn0\.880\.000\.0019\.66Orohaora1\.480\.000\.0058\.25Ososooso9\.180\.531\.0862\.09Otomi \(Family\)–0\.000\.000\.0029\.48Paamapma1\.860\.120\.2117\.59Pahariphj7\.020\.430\.8234\.09Pai Tavyterapta0\.860\.000\.001\.11Paicocasey13\.900\.821\.700\.73Palikúrplu0\.420\.000\.008\.28Panarákre0\.960\.000\.0013\.76Pantar, Westernlev0\.800\.000\.003\.90Papapanappn6\.940\.420\.8245\.16Parachiprc0\.000\.000\.000\.33Parecíspab1\.470\.000\.002\.31Parintintin \(Family\)–26\.431\.543\.01112\.03Pashai \(Family\)–0\.000\.000\.004\.74Paunakapnk6\.640\.410\.8152\.41Pechpay9\.680\.541\.104\.93Pelabxd1\.660\.000\.007\.20Penan, Easternpez0\.000\.000\.000\.07Pholaypg16\.890\.961\.98313\.63Piaroapid25\.201\.472\.86125\.55Pinghua, Southerncsp0\.000\.000\.0022\.67Pite Saamisje0\.020\.000\.0066\.35Pumapum0\.010\.000\.0020\.19Pumi, Northernpmi0\.010\.000\.0049\.18Purepechatsz0\.640\.000\.007\.55Pushtopus0\.000\.000\.007\.98Qabiaolaq0\.000\.000\.001\.28Qiang, Southernqxs3\.140\.200\.38108\.53Quetzaltepec Mixepxm0\.000\.000\.009\.34Queyuqvy0\.730\.000\.00461\.57Quichua, Loja Highlandqvj0\.000\.000\.0010\.14Q’anjob’alkjb0\.000\.000\.0056\.90Ramarma0\.000\.000\.004\.12Ratahanrth0\.480\.000\.0044\.39Rengmitca–0\.030\.000\.002\.01Ronggaror0\.060\.000\.001\.53Russianrus0\.000\.000\.007\.94Saaroasxr0\.000\.000\.001\.55Sadu–5\.540\.330\.709\.04Sakapultekoquv0\.000\.000\.0037\.11Salasaca Highland Quichuaqxl0\.560\.000\.005\.62San Martín Duraznos Mixtec–6\.210\.340\.7320\.39San Mateo del Mar Huavehuv0\.980\.000\.0015\.94Sanapanáspn2\.580\.160\.2999\.49Saparuaspr13\.830\.771\.61265\.42Sa’bansnv2\.280\.140\.275\.61Scotssco0\.000\.000\.0010\.96Seenkusos3\.360\.170\.45109\.95Sekeskj11\.700\.701\.3828\.38Sekpelelip0\.480\.000\.0029\.79Seluwasansws0\.000\.000\.0010\.10Seluwasan \(Makatian\)sws0\.630\.000\.0015\.07Serilisve0\.080\.000\.0051\.42Seruasrw0\.000\.000\.0011\.48Sgi Barajil1\.590\.090\.2010\.88Shabosbf0\.000\.000\.000\.50Shangaji–0\.000\.000\.004\.28Sharanahuamcd0\.090\.000\.000\.00Sheshx0\.100\.000\.002\.15Shehrishv1\.750\.100\.2330\.44Sheko \(Guraferdan\)she0\.490\.000\.0028\.24Sherbrobun2\.120\.110\.239\.74Sholagasle1\.630\.100\.2112\.02Shughnisgh0\.000\.000\.006\.70Shuhisxg0\.000\.000\.0072\.69Sibenco0\.380\.000\.000\.62Sihanspg1\.590\.000\.0018\.74Sikuleskh0\.000\.000\.0030\.72Simaa \(Mwenyi\)sie0\.000\.000\.003\.08Simbalismg4\.900\.270\.5584\.89Sinasina \(Kere\)sst3\.910\.240\.462\.81Singphosgp0\.000\.000\.0038\.37Sionasnn8\.690\.551\.024\.45Sirionósrq0\.000\.000\.0025\.12Sirvasbq0\.660\.000\.007\.01Skeske0\.080\.000\.0063\.73Smar\-Skad–0\.920\.000\.0062\.53Somyevkgt0\.000\.000\.000\.01Sonsorolesesov17\.261\.052\.0530\.96Soqmdc0\.880\.000\.009\.35Sotkena–0\.610\.000\.0032\.53Sowasww0\.000\.000\.005\.63Sri Lankan Portuguese Creole–17\.211\.142\.092\.23Sukisui2\.620\.150\.3333\.98Sukursyk12\.820\.751\.521\.23Sumi Nagansm0\.000\.000\.0020\.84Sumtu Chincsv1\.660\.000\.0021\.64Sunwar \(Surel\)suz0\.470\.000\.000\.01Supapyabbw0\.000\.000\.000\.03Suruísru0\.000\.000\.0051\.25Swatissw13\.670\.831\.64103\.38Syubasyw2\.090\.130\.2623\.29Tabaq–16\.540\.902\.0114\.94Tagoitag0\.000\.000\.001\.27Taltal0\.150\.000\.001\.81Tamang–2\.630\.140\.3132\.78Tangshang Naga \(Champang\)nst3\.040\.190\.3656\.77Tanimuca\-Letuamatnc0\.550\.000\.0017\.82Taromi, Uppertov0\.940\.000\.005\.79Tatuyotav0\.490\.000\.0022\.22Tehuelcheteh1\.500\.000\.0017\.56Teiwatwe0\.000\.000\.009\.86Tela\-Masbuartvm0\.230\.000\.0016\.99Tena Lowland Quichuaquw23\.231\.352\.7963\.14Teribetfr16\.410\.951\.8985\.07Teriktec2\.060\.110\.2619\.36Te’untve0\.250\.000\.0016\.02Thado Chin \(Saihriem\)tcz9\.640\.531\.10135\.20Thiek–3\.170\.180\.34130\.24Tidikelt\-Tuat Tamazight–0\.010\.000\.000\.00Tikartik0\.000\.000\.003\.17Tiniguatit5\.350\.320\.6036\.55Tobiantox0\.010\.000\.0040\.55Tok Pisintpi0\.000\.000\.00108\.85Toljic1\.350\.000\.0039\.27Tom Tatar \(Eushta\)–1\.250\.000\.0068\.85Tonsawangtnw3\.890\.220\.4416\.52Totonac, Atlequizayan–0\.000\.000\.000\.27Totonac, Filomena Mata\-Coahuitlántlp0\.000\.000\.0010\.22Totonac, Zongozotla–1\.600\.100\.203\.14Totonaco del cerro Xinolatépetltqt0\.000\.000\.007\.08To’abaitamlu0\.000\.000\.00305\.07Tshuwauhio2\.660\.150\.3111\.68Tsimanécas0\.000\.000\.0041\.99Tsúut’ínàsrs2\.450\.150\.2926\.22Tucanotuo0\.000\.000\.0025\.64Tuha–0\.160\.000\.000\.79Tujia, Southerntjs4\.100\.220\.460\.00Tulehutlu0\.000\.000\.0012\.80Tunentvu0\.000\.000\.001\.42Turoyotru2\.950\.170\.3340\.96Turungtry0\.000\.000\.0038\.04Tuvan \(Dukhan\)tyv0\.130\.000\.000\.00Tuyucatue0\.810\.000\.005\.98Tzeltaltzh69\.864\.098\.20182\.01Tîrî \(Xaa Méa\)cir0\.000\.000\.000\.73Ubanguba1\.230\.000\.0035\.04Ukaankcf2\.240\.150\.2626\.35Ukaan \(Ikaan\)kcf25\.331\.483\.05114\.81Ulwayla4\.190\.270\.496\.01Urarinaura0\.000\.000\.0015\.46Uripiv\-Wala\-Rano\-Atchin \(Atchin\)upv10\.980\.641\.3398\.03Uru\-Eu\-Wau\-Wauurz1\.300\.000\.0011\.71Uruangnirinurn8\.270\.460\.93131\.68Uspantekousp0\.000\.000\.0065\.36Vamalemkt1\.930\.120\.2569\.87Vurësmsn0\.640\.000\.0013\.39Vutevut0\.000\.000\.0021\.43V’ënen Tautnmb5\.010\.280\.632\.60Waatassn7\.760\.460\.8853\.60Wabowbb1\.220\.000\.0028\.07Waimarro0\.000\.000\.006\.68Waimahabao3\.270\.180\.394\.08Wakhiwbl0\.000\.000\.003\.68Walmanvan0\.000\.000\.003\.31Wananogvc18\.021\.062\.1822\.00Wancho Nagannp0\.000\.000\.000\.06Wardamanwrr10\.900\.671\.3072\.62Warlpiriwbp0\.000\.000\.0010\.90Waskiawsk2\.010\.130\.272\.35Wayorówyr2\.760\.170\.3332\.78Western Algerian Amazigh–5\.180\.280\.5871\.81Western Highland Chatinoctp21\.841\.332\.7679\.80Western Tarahumaratac0\.000\.000\.001\.47Wutunhuawuh0\.000\.000\.0012\.27Xicotepec de Juárez Totonactoo0\.000\.000\.0015\.68Xârâgurèaxx0\.000\.000\.000\.74Yabongybo4\.660\.300\.6063\.42Yakamayak0\.000\.000\.000\.02Yakkhaybh1\.620\.000\.001\.29Yamitao0\.000\.000\.0011\.13Yaminahuayaa1\.540\.090\.18134\.09Yan\-nhangujay1\.700\.090\.2153\.38Yaqayjaq0\.890\.000\.0014\.58Yaurjau0\.230\.000\.0011\.91Yawuno Teneyoitr0\.540\.000\.0018\.45Yeijei2\.010\.130\.2347\.89Yeresiamire0\.750\.000\.009\.24Yeretuargop0\.000\.000\.003\.58Yilyll0\.650\.000\.00174\.98Yoloxóchitl Mixtecxty66\.493\.878\.0150\.38Yopnoyut0\.110\.000\.0037\.82Yucatec Mayayua0\.540\.000\.0052\.26Yucunaycn9\.080\.521\.13133\.56Yugur, Westybe0\.130\.000\.0045\.69Yuhupyab0\.330\.000\.0060\.85Yukubenybl0\.000\.000\.004\.97Yuramts1\.430\.090\.1649\.44Yuracareyuz0\.310\.000\.0035\.68Zaghawazag1\.330\.000\.0094\.40Zapotec, Yateezty2\.300\.130\.280\.00Zauzouzal0\.290\.000\.00129\.27Zaysezay11\.480\.691\.3942\.43Zenzontepec Chatinoczn11\.960\.721\.4114\.94Zerenkelzrn2\.310\.170\.273\.85Zok–18\.351\.082\.1749\.86Äiwoonfl4\.160\.240\.5015\.12\|Gwigwj0\.000\.000\.000\.00\|\|’Amkhoe \(Sasi\)huc0\.100\.000\.0078\.73Table 7:Per\-language hours of labeled and unlabeled speech in ELARSpeech\. Labeled data are divided into train, development, and test splits, and durations below 0\.01 hours are denoted as 0\.00\. Groups of languages within a family are denoted as “Family” in parentheses, while dialects are denoted by their dialect names in parentheses \(e\.g\., Saru\)\. Multilingual subsets are excluded since their durations cannot be reliably attributed to individual languages\.Table 8:Character Error Rate \(CER\) of Whisper on the ELARSpeech test set across all 233 languages\. Mono denotes a monolingual Whisper model fine\-tuned separately for each language; Multi denotes a single multilingual Whisper model fine\-tuned on all languages\. Both are evaluated on the same test utterances\. Languages are sorted by training\-set duration and read down each column\. Data \(H\) denotes the amount of the human\-annotated training data\. Languages without an ISO 639\-3 code are marked “–”\.LanguageCodeData\(H\)ReferenceOmniVoiceLanguageCodeData\(H\)ReferenceOmniVoice\(Ground Truth\)\(Ground Truth\)CER↓\\downarrowUTMOS↑\\uparrowCER↓\\downarrowUTMOS↑\\uparrowSIM\-O↑\\uparrowCER↓\\downarrowUTMOS↑\\uparrowCER↓\\downarrowUTMOS↑\\uparrowSIM\-O↑\\uparrowMasela, Centralmxz136\.828\.392\.66012\.872\.7860\.769Bumbmv5\.510\.401\.71516\.852\.1650\.589Ju\|’hoansiktz79\.942\.932\.07246\.392\.6350\.526Karagaskim5\.429\.641\.90512\.752\.2930\.781Gorwaagow73\.362\.822\.22070\.452\.4630\.538Dzardzongke–5\.415\.702\.4047\.952\.7960\.712Tzeltaltzh69\.923\.202\.18812\.932\.6030\.692Tiniguatit5\.327\.872\.13520\.192\.2190\.693Yoloxóchitl Mixtecxty66\.58\.702\.6924\.832\.8170\.829Western Algerian Amazigh–5\.224\.242\.51112\.502\.8060\.701Isanzuisn49\.275\.032\.48280\.542\.1700\.402Mehrigdq5\.121\.862\.36020\.342\.6740\.705Emplawasemw37\.56\.862\.7035\.922\.8100\.752V’ënen Tautnmb5\.020\.982\.33413\.472\.6180\.718Dargwa \(Chirag\)dar35\.311\.282\.1493\.532\.5340\.830Animereanf5\.038\.982\.07138\.622\.3080\.655Kamasxas34\.310\.792\.12413\.432\.5780\.526Heiyohoauk4\.918\.382\.21115\.322\.3900\.659Datooga \(Asimjeeg\)tcc27\.313\.602\.49713\.342\.6260\.762Simbalismg4\.9104\.502\.05574\.132\.1720\.596Parintintin \(Family\)–26\.44\.512\.0466\.122\.4360\.684Yabongybo4\.711\.062\.6005\.532\.7650\.839Iñapariinp25\.611\.372\.51714\.972\.7520\.698Camsákbh4\.624\.122\.25813\.902\.7010\.839Ukaan \(Ikaan\)kcf25\.330\.132\.49332\.712\.6180\.668Maragusmrs4\.514\.882\.1557\.372\.5570\.728Piaroapid25\.28\.202\.48413\.142\.6690\.715Burarrabvr4\.420\.482\.1156\.122\.7380\.709Koryakkpy24\.932\.352\.26422\.732\.5510\.738Kolekl4\.415\.362\.25112\.092\.6990\.696Tena Lowland Quichuaquw23\.216\.572\.27311\.812\.6530\.717Kwakwb4\.435\.941\.98923\.532\.4530\.711Massepmvs22\.528\.811\.9937\.082\.3120\.651Fungwaula4\.320\.652\.52914\.512\.9610\.718Letilti22\.23\.972\.2344\.212\.6080\.657Hadzahts4\.236\.382\.33976\.482\.3690\.509Western Highland Chatinoctp21\.827\.602\.51717\.622\.7530\.744Notrebly4\.219\.032\.02516\.272\.6390\.606Chácobocao20\.915\.932\.45115\.622\.7430\.732Ulwayla4\.217\.332\.1698\.982\.3970\.755Matlatzinca, San Franciscomat19\.614\.602\.4158\.712\.7040\.708Äiwoonfl4\.216\.472\.58610\.002\.8290\.686Calderón Highland Quichuaqud19\.19\.592\.00310\.632\.6890\.610Kakabekke4\.119\.442\.07011\.142\.7570\.741Barambrd18\.913\.462\.4836\.622\.6650\.751Guatógta4\.114\.812\.21011\.782\.5250\.749Mulaogiu18\.688\.672\.30784\.432\.4340\.684Cora \(Family\)–4\.130\.612\.62823\.022\.8660\.774Jitotoltec–18\.613\.313\.00010\.203\.1100\.810Nheengatuyrl4\.112\.421\.90113\.212\.4080\.807Zok–18\.438\.351\.97438\.712\.4370\.696Tujia, Southerntjs4\.120\.162\.54815\.152\.9310\.773Wananogvc18\.038\.051\.87371\.972\.6530\.603Kulatpg4\.025\.602\.22412\.072\.4580\.774Dolgandlg17\.439\.982\.62334\.702\.7960\.802Kaaporurb4\.018\.442\.59410\.852\.6690\.734Kgalagadixkv17\.34\.642\.2217\.392\.3820\.678Lelepalpa3\.928\.562\.43722\.512\.6610\.736Sonsorolesesov17\.329\.272\.43372\.012\.6990\.568Enawené\-Nawéunk3\.927\.522\.30950\.582\.6070\.676Sri Lankan Portuguese Creole–17\.216\.892\.4137\.002\.7740\.785Sinasina \(Kere\)sst3\.922\.532\.67016\.242\.7480\.733Mewahang, Westernraf17\.213\.052\.29525\.192\.5860\.665Tonsawangtnw3\.920\.782\.86318\.372\.8560\.678Pholaypg16\.928\.711\.86824\.642\.2110\.596Mezquital Otomiote3\.7–––––Tabaq–16\.524\.872\.33462\.022\.7920\.584Guarayugyr3\.628\.132\.03415\.652\.6890\.708Matukar Panaumjk16\.520\.832\.39751\.082\.7730\.533Jejueojje3\.641\.101\.85825\.622\.2140\.625Teribetfr16\.434\.582\.22916\.732\.6040\.727Avatimeavn3\.540\.712\.27831\.722\.3700\.648Ibatanivb16\.220\.262\.3035\.272\.6200\.662Myenemye3\.425\.482\.04428\.572\.3880\.684Kamkdx16\.028\.012\.28237\.192\.6480\.700Kakataibo\-Kashibocbr3\.423\.162\.64615\.642\.7190\.757Okiekoki15\.318\.302\.71911\.482\.8660\.745Seenkusos3\.445\.112\.50230\.462\.7520\.758K’iche’quc15\.318\.591\.9948\.702\.1920\.751Apurinãapu3\.349\.091\.58662\.461\.8940\.591Konomalakoa15\.14\.102\.3923\.382\.7070\.752Waimahabao3\.334\.651\.95727\.692\.2970\.655Belaitbeg14\.625\.801\.94520\.872\.5200\.705Faniafni3\.214\.922\.12015\.492\.4380\.639Chimborazo Highland Quichuaqug14\.610\.422\.17016\.502\.6500\.634Kerinci \(Koto Keras\)–3\.215\.222\.4823\.732\.8130\.797Murut, Bookanbnb14\.661\.592\.41569\.762\.7710\.670Thiek–3\.215\.692\.1687\.522\.4570\.709Arapahoarp14\.423\.881\.59210\.562\.2310\.701Awiakay–3\.237\.861\.97015\.212\.4020\.587Paicocasey13\.917\.842\.4155\.712\.6660\.784Qiang, Southernqxs3\.137\.952\.47871\.292\.7080\.614Imbabura Highland Quichuaqvi13\.96\.882\.0009\.582\.5280\.613Kinnauri, Chitkulicik3\.035\.432\.07322\.472\.5500\.725Saparuaspr13\.826\.391\.82826\.062\.3070\.571Tangshang Naga \(Champang\)nst3\.021\.732\.3256\.912\.7710\.760Swatissw13\.732\.842\.44426\.372\.7490\.749Kaikekzq3\.029\.562\.38316\.952\.6100\.736E’ñapa Woromaipupbh13\.448\.452\.42521\.712\.6820\.659Turoyotru3\.012\.921\.95310\.812\.4400\.692Eveneve12\.85\.071\.79424\.062\.4860\.712Ingrianizh2\.98\.652\.4026\.282\.7940\.746Sukursyk12\.814\.842\.2376\.402\.4830\.713Northern Altai \(Teleut\)atv2\.989\.552\.33981\.862\.5150\.764Chiapas Zoque–12\.717\.112\.63411\.792\.7220\.748Kukeght2\.958\.051\.98034\.462\.2800\.636Lakurumaulxm12\.615\.642\.5287\.932\.7660\.774Binebon2\.956\.011\.98057\.702\.4750\.516Minyag, Easternemq12\.519\.842\.42227\.662\.7010\.686Kundal Shahishd2\.833\.352\.47723\.572\.9670\.824Zenzontepec Chatinoczn12\.010\.732\.3155\.942\.4740\.730Alta, Northernaqn2\.817\.922\.42315\.692\.5940\.763Sekeskj11\.728\.531\.92017\.662\.3530\.721Wayorówyr2\.819\.542\.39710\.302\.4260\.715Mafeamkv11\.610\.802\.4138\.782\.7170\.804Lavatbura\-Lamusonglbv2\.713\.752\.09012\.732\.7720\.731Zaysezay11\.523\.232\.27525\.192\.6220\.648Garifunacab2\.731\.292\.43920\.582\.7310\.696Kelabitkzi11\.58\.172\.0293\.562\.4700\.721Kandozi\-Chapracbu2\.735\.602\.13017\.592\.5950\.778Araonaaro11\.423\.431\.97616\.652\.1550\.714Tshuwauhio2\.77\.842\.72314\.542\.8280\.712Uripiv\-Wala\-Rano\-Atchin \(Atchin\)upv11\.04\.872\.1876\.102\.6430\.699Tamang–2\.625\.432\.28018\.292\.5460\.713Abuiabz11\.015\.901\.86135\.192\.4020\.577Gongdukgoe2\.615\.602\.6309\.582\.7910\.709Wardamanwrr10\.930\.262\.07616\.462\.5840\.682Sukisui2\.632\.042\.2849\.822\.5370\.703Akpesibe10\.820\.462\.11720\.242\.5140\.651Curripacokpc2\.623\.492\.27613\.462\.6160\.739Khoibu Nagankb10\.534\.212\.13420\.372\.5830\.663Sanapanáspn2\.634\.561\.88145\.662\.3040\.544Caijia–10\.028\.352\.78927\.722\.9850\.691Mormhz2\.625\.812\.44022\.862\.6340\.648Chimalapa Zoquezoh9\.814\.942\.6146\.002\.8950\.812Tsúut’ínàsrs2\.423\.232\.54421\.882\.6350\.726Pechpay9\.728\.732\.06021\.842\.5660\.721Cahuillachl2\.46\.333\.1905\.303\.0180\.807Thado Chin \(Saihriem\)tcz9\.611\.752\.11113\.812\.3050\.603Kalmyk\-Oiratxal2\.434\.361\.61717\.752\.4010\.731Mochomhc9\.625\.272\.38625\.072\.4250\.715Mororimok2\.427\.872\.07920\.122\.4130\.682Gadaba, Bodogbj9\.516\.822\.01313\.832\.4640\.668Hrusohru2\.414\.852\.73412\.953\.0100\.776Mele\-Filamxe9\.511\.372\.8047\.373\.0550\.785Arabic, Judeo\-Iraqiyhd2\.416\.182\.41410\.812\.7740\.698Batibtc9\.323\.442\.31119\.632\.6300\.662Zerenkelzrn2\.322\.062\.45611\.902\.8800\.787Bandialbqj9\.226\.212\.06516\.882\.3760\.668Zapotec, Yateezty2\.358\.172\.38563\.972\.5350\.678Ososooso9\.237\.382\.35527\.702\.7380\.783Guernésiaisnrf2\.338\.982\.11331\.862\.4970\.723Banam Bayvrt9\.132\.162\.14522\.202\.4300\.643Nalögonlz2\.320\.722\.33314\.352\.6570\.704Yucunaycn9\.127\.392\.63311\.722\.9790\.764Sa’bansnv2\.310\.652\.7297\.902\.7970\.763Eibelaail8\.910\.351\.98319\.842\.4270\.601Enxetenx2\.311\.362\.5706\.802\.6920\.715Mongolian, Peripheral \(Jirim\)mvf8\.845\.092\.20334\.772\.4550\.626Ukaankcf2\.215\.882\.44119\.652\.7120\.706Sionasnn8\.734\.992\.26722\.292\.5580\.783Lazlzz2\.27\.902\.8524\.462\.8340\.799Bastimentos Creole English–8\.655\.602\.60732\.902\.9840\.594Nese–2\.230\.502\.33213\.872\.6150\.717Analanm8\.537\.232\.03620\.722\.6230\.594Kichwa de Chibuleo–2\.29\.571\.5314\.111\.9960\.719Nkaminkq8\.428\.921\.88121\.352\.5170\.641Mungbam \(Ngun\)mij2\.288\.312\.00149\.642\.5050\.596Uruangnirinurn8\.38\.552\.29212\.102\.6020\.694Sherbrobun2\.1103\.072\.24094\.362\.6220\.675Huillichehuh8\.28\.242\.3918\.122\.7850\.716Bihibh2\.118\.042\.3209\.922\.6180\.778Kpashampbn8\.124\.442\.86519\.383\.0860\.659Syubasyw2\.126\.062\.27419\.582\.5720\.673Minyag, Westernwmg8\.124\.562\.53812\.152\.7770\.747Datoogatcc2\.116\.371\.81911\.512\.4490\.683Nsiteubko7\.931\.672\.02823\.692\.3480\.529Teriktec2\.134\.022\.20728\.072\.7010\.603Baniwabwi7\.820\.182\.40210\.492\.7750\.761Jowshaqani–2\.020\.982\.47710\.192\.8270\.777Gwenogwe7\.850\.312\.32544\.952\.5490\.717Waskiawsk2\.05\.992\.2033\.832\.6120\.797Jawejaz7\.814\.592\.2049\.722\.5390\.735Yeijei2\.048\.782\.47939\.942\.6540\.671Waatassn7\.812\.232\.17614\.782\.5070\.657Kalmyk\-Oirat \(Durvud\)xal1\.936\.051\.71522\.892\.2890\.632Dawoodidmk7\.622\.871\.93924\.262\.3260\.851Vamalemkt1\.930\.942\.15621\.832\.4840\.678Makurápmpu7\.317\.352\.17110\.962\.4270\.759Ayoreoayo1\.950\.450\.29122\.731\.6040\.329Dâwkwa7\.281\.682\.09665\.312\.3540\.729Paamapma1\.971\.992\.36072\.382\.4750\.507Pahariphj7\.026\.682\.16029\.042\.4500\.692Awutuafu1\.836\.162\.73124\.532\.9430\.763Papapanappn6\.930\.912\.35425\.752\.7920\.716Maemme1\.819\.352\.51411\.412\.6700\.745Arotei6\.812\.032\.2547\.712\.4010\.819Kangjiakxs1\.824\.432\.38912\.282\.7440\.801Paunakapnk6\.626\.502\.30011\.832\.8730\.658Shehrishv1\.823\.552\.56523\.022\.8020\.726Jabutíjbt6\.531\.512\.36523\.172\.4490\.691Ersuers1\.715\.902\.54713\.212\.8010\.718Blablangablp6\.416\.782\.49723\.382\.7440\.740Bareinbva1\.723\.582\.71016\.672\.8690\.725Muyu–6\.323\.582\.26915\.842\.5130\.677Mbyá Guaranígun1\.726\.182\.42820\.012\.7920\.815San Martín Duraznos Mixtec–6\.247\.292\.25731\.722\.6370\.735Yan\-nhangujay1\.713\.072\.5416\.492\.8620\.760Burarra \(Gun\-nartpa\)bvr6\.116\.292\.33814\.402\.6390\.671Angaitéaqt1\.616\.792\.17913\.062\.4670\.769Gargarnaye–6\.029\.972\.19318\.492\.5090\.717Sholagasle1\.610\.961\.94817\.312\.0190\.711Cicipuawc6\.032\.222\.57129\.942\.7210\.677Chiquitano \(Bésiro\)cax1\.66\.352\.30576\.022\.6150\.574Gulf Arabicafb6\.022\.282\.14713\.492\.5280\.748Júmajua1\.61\.711\.7135\.152\.1080\.773Highland Popolucapoi5\.817\.822\.41511\.372\.5560\.756Totonac, Zongozotla–1\.6–––––Chiniafi5\.766\.232\.11173\.652\.3000\.472Sgi Barajil1\.614\.382\.20110\.672\.2910\.722Cofáncon5\.726\.481\.92023\.332\.3580\.686Brembuq1\.615\.402\.14911\.372\.6150\.765Bunakbfn5\.668\.182\.47154\.482\.7840\.666Cabiyarícbb1\.649\.771\.96344\.802\.0710\.705Ghanonggaghn5\.68\.822\.4754\.652\.7220\.778Ibibioibb1\.655\.321\.69939\.762\.6790\.675Sadu–5\.5124\.352\.426112\.492\.6210\.556Yaminahuayaa1\.52\.942\.2132\.972\.3970\.714Ambelwgo5\.522\.842\.06513\.382\.5790\.718Yuramts1\.420\.622\.11315\.522\.3550\.677Emaemmw5\.514\.232\.41611\.522\.7100\.797Bongubpu1\.418\.322\.3857\.402\.5360\.749Akhaahk5\.514\.921\.99611\.942\.2130\.785Table 9:CER, UTMOS, and speaker similarity \(SIM\-O\) of the multilingual OmniVoice model on the ELARSpeech test set\. UTMOS and OmniVoice results for Mezquital Otomi and Totonac, Zongozotla are unavailable because no test samples remain after SNR/C50 filtering\. Languages are sorted by training\-set duration and read down each column\. Data \(H\) denotes the hours of human\-annotated training data\. Languages without an ISO 639\-3 code are marked “–”\. Reference CER and UTMOS were measured with a monolingual Whisper model and UTMOS\-v2, respectively, on ground\-truth speech\.