Meta’s Omnilingual ASR transcribes more than 1,600 languages, including over 500 that no speech recognition system had ever served (Meta FAIR, Omnilingual ASR 2025), yet that is still under a quarter of the 7,170 living languages in Ethnologue’s 2026 edition. Two years earlier the reference point was Whisper, trained on 680,000 hours with just 117,000 hours spread across 96 non-English languages (OpenAI, Whisper paper 2022). The biggest open crowd-sourced corpus, Mozilla Common Voice, now spans 289 languages and 38,932 hours (Common Voice 24.0, 2025), and the fastest-growing products, from live translation to dictation, all sit on top of these foundation models, as our speech-to-text statistics roundup also tracks. We aggregated data from Meta FAIR, OpenAI, Google Research, Ethnologue, UNESCO, Mozilla, NVIDIA, Hugging Face, PNAS, MarketsandMarkets and other primary sources listed in the methodology.
TL;DR
- Omnilingual ASR covers 1,600+ languages, 500+ of them never served by ASR before (Meta FAIR, Omnilingual ASR 2025).
- Zero-shot in-context learning could extend Omnilingual ASR to more than 5,400 languages (Meta, November 2025).
- 7,170 living languages exist, a net gain of 11 in one year (Ethnologue, 29th edition 2026).
- 40% of the world’s population does not access education in a language they understand (UNESCO GEM Report 2025).
- Meta’s MMS identifies 4,017 spoken languages and transcribes 1,107 (Meta, MMS paper 2023).
- Google’s USM trained on 12 million hours of speech across 300+ languages (Google Research 2023).
- Whisper large-v3 trained on about 5 million hours, 4 million of them pseudo-labeled (OpenAI, Nov 2023).
- Whisper word error rate halves for every 16x increase in training data per language (OpenAI, Whisper paper 2022).
- Common Voice 24.0: 38,932 hours across 289 languages (Mozilla, Dec 2025).
- Character error rate below 10 for 78% of Omnilingual ASR’s languages (Meta, Nov 2025).
- Gemini 3.5 Live Translate supports 70+ languages and 2,000+ combinations in one Meet call (Google, June 2026).
- Black speakers saw an average WER of 0.35 versus 0.19 for white speakers across five commercial systems (PNAS, 2020).
1. The Language Gap: 7,170 Languages, a Few Hundred Served
The denominator matters. Ethnologue’s 29th edition counts 7,170 living languages, a net increase of 11 over the prior year after more than 8,000 database updates, with eight languages reclassified from extinct to living (Ethnologue, Welcome to the 29th edition, February 2026). Before 2023, the MMS team noted, mainstream speech systems handled only around 100 of them. Even after the jump to 1,600+, roughly three in four living languages still have no off-the-shelf speech recognizer.
The gap is not academic. UNESCO’s Global Education Monitoring team reports that 40% of the global population does not access education in a language they understand, a share that reaches 90% in some low- and middle-income countries (UNESCO GEM Report, Languages Matter). Speech technology that skips those languages also skips the people for whom voice interfaces would matter most, a theme our endangered languages statistics cover from the preservation side.
| Metric | Value | Source |
|---|---|---|
| Living languages, 2026 | 7,170 (net +11 in one year) | Ethnologue, 29th edition 2026 |
| Living languages, 2025 | 7,159 | Ethnologue, 28th edition 2025 |
| Endangered languages | 3,193 (45% of living languages) | Ethnologue, 28th edition 2025 |
| ’Dying’ languages | 1,030 (14%) | Ethnologue, 28th edition 2025 |
| Population without education in a language they understand | 40% (up to 90% in some LMICs) | UNESCO GEM Report 2025 |
| Languages covered by typical speech systems before MMS | About 100 | Meta, MMS paper 2023 |
| Languages never before served by ASR, added by Omnilingual ASR | More than 500 | Meta FAIR, Omnilingual ASR 2025 |
Context note: the endangered and dying counts are from the 2025 edition as summarized in public reporting; Ethnologue’s full figures sit behind its subscription. UNESCO pairs the 40% figure with large learning penalties: in Honduras, 94% of grade 6 students who spoke the language of instruction achieved reading basics versus 62% who did not.
2. Coverage Race: How Many Languages Each Model Handles
Language counts jumped by an order of magnitude in three years. Meta’s MMS identifies 4,017 languages and transcribes and synthesizes 1,107, built largely from New Testament recordings, with pre-trained models covering 1,406 languages (MMS: Scaling Speech Technology to 1,000+ Languages). Omnilingual ASR then pushed transcription past 1,600 and, through zero-shot in-context learning from a few paired examples, claims potential reach of more than 5,400 languages (Meta AI, Omnilingual ASR announcement).
Commercial products are far more conservative, and the gap tells you where the quality bar sits. Google Cloud’s Chirp 3 lists 24 generally available transcription languages and 76 in preview (Google Cloud, Chirp 3 documentation), while research models claim hundreds. Speech translation sits in between: Meta’s SeamlessM4T takes speech in 101 languages but speaks in only 36 (Nature, Joint speech and text machine translation for up to 100 languages), a reminder that generating speech is harder than recognizing it.
| Metric | Value | Source |
|---|---|---|
| Omnilingual ASR transcription | 1,600+ languages | Meta FAIR, Nov 2025 |
| Omnilingual ASR potential via zero-shot learning | 5,400+ languages | Meta, Nov 2025 |
| MMS language identification / ASR and TTS / pre-training | 4,017 / 1,107 / 1,406 languages | Meta, MMS paper 2023 |
| Google USM pre-training coverage | 300+ languages | Google Research, March 2023 |
| Whisper large-v3 | 99 languages | Hugging Face Open ASR Leaderboard 2025 |
| SeamlessM4T speech-to-speech translation | 101 input, 36 output languages | Meta, Nature 2025 |
| Google Chirp 3 transcription | 24 GA + 76 preview languages | Google Cloud documentation 2026 |
| Meta OmniSONAR speech embeddings | 177 spoken languages in a 200-language semantic space | Meta, OmniSONAR paper 2026 |
Real-time translation has crossed into mass-market apps. Gemini 3.5 Live Translate detects 70+ languages and supports more than 2,000 language combinations in one Google Meet session, up from five languages (Google, Gemini 3.5 Live Translate, June 2026). Google says people translate around 1 trillion words a month across Translate, Search, Lens and Circle to Search (Google, August 2025). More on device-level adoption is in our speech-to-speech translation statistics.
| Metric | Value | Source |
|---|---|---|
| Gemini 3.5 Live Translate languages | 70+ | Google, June 2026 |
| Language combinations in one Meet call | 2,000+ (previously 5 languages) | Google, June 2026 |
| Google Translate live conversation languages | 70+ | Google, August 2025 |
| Words translated per month across Google products | About 1 trillion | Google, August 2025 |
| Mistral Voxtral open-weight model sizes | 24B and 3B parameters | Mistral AI, July 2025 |
| Voxtral maximum audio length | 30 minutes transcription, 40 minutes understanding | Mistral AI, July 2025 |
3. Training Data Scale: From 680,000 Hours to 12 Million
The data budget behind these models grew far faster than language counts. Google’s Universal Speech Model was trained on 12 million hours of speech and 28 billion sentences of text, with 2 billion parameters (Google Research, USM), roughly 18 times Whisper’s original corpus. OpenAI itself went from 680,000 hours to about 5 million for Whisper large-v3, of which 4 million were pseudo-labeled by the previous model (OpenAI, large-v3 release notes).
The distribution inside these corpora is the real story. Whisper’s original 680,000 hours included 117,000 hours across 96 non-English languages and 125,000 hours of translation into English (Whisper paper). By subtraction (680,000 minus 117,000 minus 125,000), about 438,000 hours, or roughly 64%, was English transcription. Meta’s MMS labeled set averaged only 32 hours per language, and Omnilingual ASR’s released data averages about 10 transcribed hours per language. Massively multilingual models are built on a long tail measured in hours, not thousands of hours.
| Metric | Value | Source |
|---|---|---|
| Whisper training audio | 680,000 hours | OpenAI, Whisper paper 2022 |
| Whisper non-English recognition data | 117,000 hours across 96 languages | OpenAI, Whisper paper 2022 |
| Whisper X-to-English translation data | 125,000 hours | OpenAI, Whisper paper 2022 |
| Whisper English transcription (derived) | About 438,000 hours (about 64%) | Calculated from OpenAI figures |
| Whisper large-v3 training data | 1M weakly labeled + 4M pseudo-labeled hours | OpenAI, Nov 2023 |
| Google USM | 12M hours speech, 28B text sentences, 2B parameters | Google Research, 2023 |
| MMS self-supervised pre-training | About 500,000 hours across 1,400+ languages | Meta, May 2023 |
| MMS labeled data per language | 32 hours on average | Meta, May 2023 |
Meta’s newest stack scales both ends. Omnilingual ASR pre-trained a 7-billion-parameter wav2vec 2.0 encoder on 3.84 million hours across 1,239 languages, plus 460,000 hours with no language label, and its supervised AllASR corpus holds 120,710 hours across 1,690 languages (Omnilingual ASR paper). The smallest released variant has 300 million parameters, aimed at low-power devices.
| Metric | Value | Source |
|---|---|---|
| Omnilingual ASR unlabeled pre-training | 3.84M hours, 1,239 languages | Meta FAIR, Omnilingual ASR paper 2025 |
| Additional unlabeled speech without language ID | 460,000 hours | Meta FAIR, 2025 |
| AllASR labeled corpus | 120,710 hours, 1,690 languages | Meta FAIR, 2025 |
| Average transcribed speech per language (released data) | About 10 hours | Meta FAIR, 2025 |
| Largest / smallest model | 7B / 300M parameters | Meta FAIR, 2025 |
4. Open Datasets and Benchmarks Behind the Models
Open data is where long-tail coverage gets built. Mozilla Common Voice Scripted Speech 24.0 reached 38,932 hours across 289 languages, 25,886 of them validated (Mozilla Discourse release post, December 2025). The bigger leap came one release earlier, when Common Voice 23.0 added 149 languages and more than doubled its count to 286 (Mozilla Data Collective). Hours grew only about 8% between those releases, so new languages arrive with very little audio each.
Evaluation remains far narrower than training. FLEURS, the standard multilingual benchmark, covers 102 languages with about 12 hours of speech per language (FLEURS paper), and the Hugging Face Open ASR Leaderboard’s multilingual track benchmarks just five languages across 60+ models from 18 organizations (Hugging Face, Open ASR Leaderboard). Claims of ‘1,600 languages’ are therefore checked on a fraction of that number.
| Metric | Value | Source |
|---|---|---|
| Common Voice 24.0 total / validated hours | 38,932 / 25,886 | Mozilla, Dec 2025 |
| Common Voice 24.0 languages | 289 | Mozilla, Dec 2025 |
| Common Voice 23.0 languages added | 149 (total 286) | Mozilla Data Collective, Sept 2025 |
| Common Voice 23.0 spontaneous speech | 357 hours across 51 languages | Mozilla Data Collective, Sept 2025 |
| Spontaneous Speech 2.0 datasets | 62 | Mozilla, Dec 2025 |
| FLEURS benchmark | 102 languages, about 12 hours each | Conneau et al., FLEURS 2022 |
| Omnilingual ASR Corpus | 350 underserved languages (CC-BY) | Meta, Nov 2025 |
Europe got its own large open corpus. NVIDIA’s Granary holds about 1 million hours: nearly 650,000 for recognition and more than 350,000 for translation, across 25 European languages including Croatian, Estonian and Maltese (NVIDIA blog, August 2025). The accompanying paper, accepted at Interspeech 2025, reports models reach similar performance with roughly 50% less data than with prior datasets (Granary paper).
| Metric | Value | Source |
|---|---|---|
| Granary total audio | About 1 million hours | NVIDIA, Aug 2025 |
| Granary recognition / translation split | About 650,000 / 350,000+ hours | NVIDIA, Aug 2025 |
| Granary languages | 25 European | NVIDIA, Aug 2025 |
| Data needed vs prior datasets for same accuracy | About 50% less | Granary paper, Interspeech 2025 |
| Open ASR Leaderboard coverage | 60+ models, 18 organizations, 11 datasets | Hugging Face, Nov 2025 |
| ASR models hosted on the Hugging Face Hub | About 27,000 | Hugging Face, Nov 2025 |
| Languages in the leaderboard’s multilingual track | 5 | Hugging Face, Nov 2025 |
5. Accuracy: Data per Language Beats Model Size
The most useful number in the field is a scaling law. Whisper’s authors found a squared correlation of 0.83 between log word error rate and log training hours per language, with WER halving for every 16x increase in data (OpenAI, Whisper paper). The same relationship was only 0.24 for speech translation. For a language with 10 hours of audio, the practical path to usable accuracy is more data, not a bigger model.
Coverage gains have come with surprisingly small accuracy costs. Meta reported that scaling MMS from 61 to 1,107 languages raised character error rate by only about 0.4%, while MMS halved Whisper’s word error rate on 54 FLEURS languages (Meta AI, MMS announcement). Omnilingual ASR reaches a character error rate below 10 for 78% of its 1,600+ languages. Self-reported, but each claim is tied to a public benchmark.
| Metric | Value | Source |
|---|---|---|
| Correlation, log WER vs log training hours (ASR) | r squared = 0.83 | OpenAI, Whisper paper 2022 |
| Same correlation for speech translation | r squared = 0.24 | OpenAI, Whisper paper 2022 |
| WER improvement per 16x more data | Halves | OpenAI, Whisper paper 2022 |
| MMS vs Whisper on 54 FLEURS languages | About half the WER | Meta, MMS 2023 |
| CER increase scaling MMS from 61 to 1,107 languages | About 0.4% | Meta, May 2023 |
| Omnilingual ASR languages with CER below 10 | 78% | Meta, Nov 2025 |
| Whisper large-v3 error reduction vs large-v2 | 10% to 20% | OpenAI, Nov 2023 |
Google’s numbers point the same way. USM posted 32.7% relative lower WER than Whisper on 18 comparable languages and under 30% average WER across 73 YouTube caption languages, with 65.8% lower relative WER on FLEURS without in-domain data. Efficiency is now a second race: NVIDIA’s billion-parameter Canary-1b-v2 claims quality comparable to models 3x larger at up to 10x faster inference, and on the long-form leaderboard Parakeet CTC posted an RTFx of 2,793.75 versus 68.56 for Whisper large-v3 (Hugging Face, 2025). For real-time voice tools, including the kind behind live voice cloning software, that throughput gap often matters more than a point of WER.
| Metric | Value | Source |
|---|---|---|
| USM WER vs Whisper, 18 languages | 32.7% relatively lower | Google Research, 2023 |
| USM average WER, 73 YouTube languages | Under 30% | Google Research, 2023 |
| USM relative WER reduction on FLEURS (no in-domain data) | 65.8% | Google Research, 2023 |
| SeamlessM4T v2 vs cascaded systems | Up to 8% (speech-to-text) and 23% (speech-to-speech) higher BLEU | Meta, Nature 2025 |
| OmniSONAR speech-to-text translation, zero-shot | 97% of SeamlessM4T quality | Meta, OmniSONAR 2026 |
| Canary-1b-v2 vs larger models | Comparable quality to 3x larger, up to 10x faster | NVIDIA, Aug 2025 |
| Long-form throughput (RTFx): Parakeet CTC vs Whisper large-v3 | 2,793.75 vs 68.56 | Hugging Face Open ASR Leaderboard 2025 |
Outlier note: the OmniSONAR paper (arXiv, March 2026) also reports 43% lower speech similarity-search error than prior models, a sign that shared speech-text embedding spaces are becoming a building block in their own right.
6. Accent Bias, Access and the Commercial Stakes
Language coverage hides a second gap inside languages that are ‘supported’. Across five commercial systems from Amazon, Apple, Google, IBM and Microsoft, the average WER was 0.35 for Black speakers versus 0.19 for white speakers (Koenecke et al., PNAS 2020), based on interviews with 73 Black and 42 white speakers. It is older than three years and is shown as the most recent available audit of its kind. Newer work reaches similar conclusions: AfriSpeech-MultiBench, covering 100+ African English accents in 10+ countries, found proprietary models accurate on clean speech but varying significantly by country and domain, with hallucinations still a problem for most state-of-the-art models (arXiv, Nov 2025).
The commercial stakes keep rising, though market sizing is inconsistent. MarketsandMarkets projects speech and voice recognition to grow from $9.66 billion in 2025 to $23.11 billion by 2030, a 19.1% CAGR (MarketsandMarkets press release), while Grand View Research puts the market at $20.2 billion in 2023 heading to $53.7 billion by 2030 at 14.6% (Grand View Research). Both firms flag Asia Pacific and emerging markets as growth drivers, which are exactly the regions where language coverage is thinnest. See also our machine translation statistics for the text side of the same market.
| Metric | Value | Source |
|---|---|---|
| Average WER, Black vs white speakers (5 systems) | 0.35 vs 0.19 | Koenecke et al., PNAS 2020 |
| Speakers in the PNAS audit | 73 Black, 42 white | Koenecke et al., PNAS 2020 |
| African English accents benchmarked | 100+ across 10+ countries | AfriSpeech-MultiBench, 2025 |
| Grade 5 reading basics, Côte d’Ivoire: home language = test language vs not | 55% vs 25% | UNESCO GEM Report 2025 |
| Speech and voice recognition market, 2025 to 2030 | $9.66B to $23.11B (19.1% CAGR) | MarketsandMarkets, Aug 2025 |
| US speech and voice recognition market, 2025 to 2030 | $3.13B to $7.70B (19.7% CAGR) | MarketsandMarkets, 2025 |
| Voice and speech recognition market, 2023 to 2030 | $20.2B to $53.7B (14.6% CAGR) | Grand View Research, 2024 |
| North America revenue share | 30.8% (2023) | Grand View Research, 2024 |
Summary: Multilingual Speech Foundation Models by the Numbers
| Metric | Value | Source |
|---|---|---|
| Living languages | 7,170 | Ethnologue, 29th edition 2026 |
| Omnilingual ASR coverage | 1,600+ languages (500+ never served before) | Meta FAIR, 2025 |
| Omnilingual ASR zero-shot potential | 5,400+ languages | Meta, Nov 2025 |
| MMS language identification | 4,017 languages | Meta, 2023 |
| Whisper large-v3 coverage | 99 languages | Hugging Face, 2025 |
| SeamlessM4T speech input / speech output | 101 / 36 languages | Meta, Nature 2025 |
| Gemini 3.5 Live Translate | 70+ languages, 2,000+ combinations | Google, June 2026 |
| Whisper training data | 680,000 hours (117,000 non-English) | OpenAI, 2022 |
| Whisper large-v3 training data | About 5 million hours | OpenAI, 2023 |
| Google USM training data | 12 million hours, 300+ languages | Google Research, 2023 |
| Omnilingual ASR pre-training | 3.84M hours, 1,239 languages | Meta FAIR, 2025 |
| Common Voice 24.0 | 38,932 hours, 289 languages | Mozilla, Dec 2025 |
| NVIDIA Granary | About 1M hours, 25 languages | NVIDIA, Aug 2025 |
| FLEURS benchmark | 102 languages, about 12 hours each | FLEURS, 2022 |
| WER vs data | Halves per 16x more data | OpenAI, 2022 |
| Omnilingual ASR languages with CER below 10 | 78% | Meta, Nov 2025 |
| MMS vs Whisper WER, 54 languages | About half | Meta, 2023 |
| Population without education in a language they understand | 40% | UNESCO, 2025 |
| WER gap, Black vs white speakers | 0.35 vs 0.19 | PNAS, 2020 |
| Speech and voice recognition market, 2030 | $23.11B (MarketsandMarkets) vs $53.7B (Grand View) | MarketsandMarkets 2025; Grand View Research 2024 |
Methodology and Sources
Every figure was traced to the organization that produced it, preferring papers, official release notes and documentation over press coverage. Vendor accuracy claims (Meta, Google, OpenAI, NVIDIA, Mistral) are self-reported on public benchmarks and are labeled with their source. Commercial claims we could not tie to a published number, such as unquantified ‘lowest WER’ marketing statements, were excluded.
- Ethnologue: Welcome to the 29th edition (Feb 2026)
- UNESCO GEM Report: 40% don’t access education in a language they understand (2025)
- Meta: MMS announcement (May 2023), MMS paper, Omnilingual ASR announcement (Nov 2025), Omnilingual ASR paper, OmniSONAR paper (2026)
- Meta Seamless Communication team: Joint speech and text machine translation for up to 100 languages (Nature, 2025)
- OpenAI: Robust Speech Recognition via Large-Scale Weak Supervision (2022), Whisper large-v3 release (Nov 2023)
- Google: Universal Speech Model (March 2023), Chirp 3 documentation, Gemini 3.5 Live Translate (June 2026), Translate live translation (Aug 2025)
- Mozilla: Common Voice 24.0 release (Dec 2025), Common Voice 23.0 release (Sept 2025)
- Conneau et al.: FLEURS (2022)
- NVIDIA: Granary, Canary and Parakeet release (Aug 2025), Granary paper (Interspeech 2025)
- Hugging Face: Open ASR Leaderboard trends (Nov 2025)
- Mistral AI: Voxtral (July 2025)
- Koenecke et al.: Racial disparities in automated speech recognition (PNAS, 2020)
- AfriSpeech-MultiBench: arXiv 2511.14255 (IJCNLP-AACL 2025)
- MarketsandMarkets: Speech and Voice Recognition Market press release (Aug 2025)
- Grand View Research: Voice and Speech Recognition Market Report (2024)
- Data watch: Market sizing conflicts sharply. MarketsandMarkets ($23.11B by 2030) and Grand View Research ($53.7B by 2030) use different scopes and base years, so the numbers should not be averaged or compared as a trend. Several foundation-model figures date from 2022 to 2023 (Whisper, USM, MMS, FLEURS) and are the most recent available disclosures for those models; Google has not published comparable training-data totals for its Gemini audio models. The English share of Whisper’s data (about 438,000 hours, 64%) is our subtraction from OpenAI’s published totals, not a figure OpenAI states in the abstract. Ethnologue’s 2025 endangered and dying counts come from public summaries of the 28th edition because the full database is paywalled. The PNAS bias audit (2020) predates current foundation models and is shown as most recent available data; no equally large cross-vendor audit has been published since. Language counts are not comparable across vendors: some count locales, some count languages, and ‘preview’ coverage is not production quality.
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when Mozilla publishes Common Voice 25.0 and Ethnologue releases its 30th edition in February 2027.