The speech recognition training data behind OpenAI’s Whisper held 438,218 hours of English but just 0.1 hours each of Lao, Sundanese and Burmese (OpenAI, Whisper paper 2022). That gap is the long tail of speech recognition in one line, set against 7,170 languages in use worldwide (Ethnologue, How Many Languages 2026). The floor for useful accuracy looks lower than that gap suggests: one 2025 study reached a word error rate below 13% with as little as 50 hours of training data (Akera et al., Kinyarwanda and Kikuyu data-scaling study 2025). Crowdsourcing is filling part of the gap, with Mozilla Common Voice at 29,295 validated hours (Mozilla, Common Voice cv-dataset repository 2026). Accuracy gaps across accents inside major languages are covered in our speech recognition accent bias statistics; this roundup looks at whole languages. We aggregated data from Ethnologue, the OpenAI Whisper paper, Meta’s MMS and Omnilingual ASR papers, Google’s USM and FLEURS papers, Mozilla Common Voice, Project Vaani, IndicVoices, NaijaVoices, African Next Voices, and recent African-language ASR benchmarks.
TL;DR
- 7,170 languages are in use today, and about 44% are endangered, often with fewer than 1,000 users (Ethnologue, How Many Languages 2026).
- Only 195 of 546 low-resource languages (36%) reached a character error rate below 10 in Meta’s 7B model, against 236 of 249 high-resource languages (Meta FAIR, Omnilingual ASR paper 2025).
- 65% of Whisper’s training audio was English speech recognition, 438,218 hours, while 96 other languages shared 117,113 hours (OpenAI, Whisper paper 2022).
- A Kinyarwanda model reached a word error rate below 13% with just 50 hours of training data (Akera et al., data-scaling study 2025).
- The best Swahili model fell from 24.81% WER with 1 hour of data to 11.36% with 400 hours (Nahabwe et al., Benchmarking ASR Models for African Languages 2025).
- Mozilla Common Voice Scripted Speech holds 42,593 hours, 29,295 of them validated, in release 27.0 (Mozilla, Common Voice cv-dataset repository 2026).
- Project Vaani has recorded 31,278.37 hours of Indian speech, but only about 6.8% is transcribed (IISc and ARTPARK, Project Vaani dashboard 2026).
- African Next Voices recorded 9,000+ hours in 18 languages across Kenya, Nigeria and South Africa (University of Pretoria, African Next Voices news 2025).
- Fine-tuning on the 1,800-hour NaijaVoices corpus improved Whisper’s average word error rate by 75.86% on Igbo, Hausa and Yoruba (NaijaVoices dataset paper, Interspeech 2025).
- Google’s Universal Speech Model beat Whisper by 66% relative WER on 62 FLEURS languages with a labeled set about one seventh the size (Google, USM paper 2023).
- IndicVoices holds 7,348 hours from 16,237 speakers across 22 Indian languages (IndicVoices dataset paper 2024).
1. How Many of the World’s Languages Have Any Speech Recognition
The denominator is the story. Ethnologue counts 7,170 languages in use today, and just 20 of them, 0.3% of the total, are the native languages of more than 3.7 billion people (Ethnologue, How Many Languages 2026). Speech recognition followed that concentration: as late as 2023, Google’s Universal Speech Model covered 100+ languages.
Even the broadest open system today, at 1,600+ languages, reaches only about a fifth of Ethnologue’s count. Language loss is the other side of this gap, covered in our endangered languages statistics. Africa alone accounts for more than 2,000 languages, many of them lacking transcribed speech data.
| Metric | Value | Source |
|---|---|---|
| Languages in use worldwide | 7,170 | Ethnologue, How Many Languages 2026 |
| Share of living languages that are endangered | About 44%, often with fewer than 1,000 users remaining | Ethnologue, How Many Languages 2026 |
| Native speakers of the 20 largest languages | Over 3.7 billion people, nearly half the world population, from 0.3% of all languages | Ethnologue, How Many Languages 2026 |
| Languages spoken across Africa | More than 2,000, many lacking transcribed speech data | Nahabwe et al., Benchmarking ASR Models for African Languages 2025 |
| Languages covered by Google’s Universal Speech Model for speech recognition | 100+ | Google, USM paper 2023 |
| Languages of the 8th Schedule of India’s Constitution covered by IndicASR, which its authors call the first model to support all of them | 22 | IndicVoices dataset paper 2024 |
| Languages covered by the broadest open ASR system | 1,600+ (over 500 never served by any ASR system before), about 22% of the 7,170 in use (1,600 / 7,170) | Meta FAIR, Omnilingual ASR paper 2025; Ethnologue 2026 |
Context: that 1,600+ system deliberately leaves out roughly 150 signed languages, constructed languages and languages without an established writing system, so part of the uncovered total is outside the reach of speech recognition that outputs text. Most recent available data for USM: 2023.
2. Training Hours: English vs the Long Tail
Training data for speech recognition is not a distribution, it is a cliff. Whisper’s English speech recognition data (438,218 hours) was roughly 81,000 times larger than its Swahili data (5.4 hours) (438,218 / 5.4 = 81,151), and Lao, Sundanese and Burmese got 0.1 hours each. Even against Chinese, the best-supplied non-English language at 23,446 hours, English had about 19 times as much audio (438,218 / 23,446 = 18.7).
Newer systems sort languages into resource buckets, and the boundaries are telling: a language counts as high-resource (well supplied with training data) above just 50 hours. Purpose-built datasets can move a language across that line: IndicVoices reports a median of 73 transcribed hours per language across its 22 languages (IndicVoices dataset paper 2024).
| Metric | Value | Source |
|---|---|---|
| Whisper training audio: English speech recognition vs 96 other languages | 438,218 hours (65%) vs 117,113 hours (17%) | OpenAI, Whisper paper 2022 |
| Whisper speech recognition hours per language, largest vs smallest | Chinese 23,446; German 13,344; Spanish 11,100 vs Swahili 5.4; Lao, Sundanese and Burmese 0.1 each | OpenAI, Whisper paper 2022 |
| FLEURS benchmark languages with no Whisper training data at all | 20 of 102 | OpenAI, Whisper paper 2022 |
| Omnilingual ASR languages by labeled training data: more than 50 hours / 10 to 50 hours / under 10 hours | 249 / 881 / 546 languages | Meta FAIR, Omnilingual ASR paper 2025 |
| FLEURS evaluation data per language | About 12 hours of speech in each of 102 languages | Google, FLEURS paper 2022 |
| IndicVoices transcribed audio | 1,639 hours, median 73 hours per language | IndicVoices dataset paper 2024 |
Most recent available data for Whisper: 2022. The hours above are speech recognition data only, separate from Whisper’s translation audio. We found no equivalent per-language breakdown for later Whisper versions.
3. Error Rates: Low-Resource vs High-Resource Languages
Word error rate (WER) is the share of words a system gets wrong, counting substitutions, deletions and insertions; because inserted words count, it can exceed 100% when a model produces more wrong words than the reference contains. Whisper large-v2 scored 4.2% WER in English but 121.0% in Shona on the FLEURS benchmark, so for Shona the errors outnumbered the words actually spoken (OpenAI, Whisper paper 2022). One known way a model gets there is by writing fluent text nobody said, the failure mode tracked in our speech recognition hallucination statistics.
Newer models still leave the gap in place. In the table below, character error rate (CER, the same idea counted per letter rather than per word) stayed under 10 for about 95% of languages with 10 or more hours of data (1,077 of 1,130), but for just over a third of the under-10-hour group. Inside one model, data volume alone separates those outcomes.
| Metric | Value | Source |
|---|---|---|
| Whisper large-v2 word error rate on FLEURS: English / Spanish / Swahili | 4.2% / 3.0% / 39.3% | OpenAI, Whisper paper 2022 |
| Whisper large-v2 word error rate on FLEURS: Yoruba / Somali / Shona / Sindhi | 94.8% / 102.9% / 121.0% / 156.5% | OpenAI, Whisper paper 2022 |
| Omnilingual 7B LLM-ASR languages with character error rate below 10, by resource level: high / mid / low | 236 of 249 / 841 of 881 / 195 of 546 (the paper rounds the low-resource share to 36%) | Meta FAIR, Omnilingual ASR paper 2025 |
| Omnilingual 7B LLM-ASR mean character error rate: high-resource vs low-resource languages | 3.13 vs 18.0 | Meta FAIR, Omnilingual ASR paper 2025 |
Outlier note: the Whisper rows come from the large-v2 table in the 2022 paper, the most recent available per-language table of this kind. WER and CER are different measures, so the Whisper and Omnilingual rows cannot be compared directly.
4. How Much Data Closes the Gap
The encouraging finding of 2025 is that the threshold for usable recognition is measured in tens of hours, not thousands. A Kinyarwanda model reached practical accuracy, a word error rate below 13%, with as little as 50 hours of training data, and dropped below 10% by 200 hours (Akera et al., data-scaling study 2025). If that holds for other languages, one funded recording campaign is enough for a first usable model.
Data quality counts too. The same study traced 38.6% of high-error cases to noisy ground-truth transcriptions, meaning a share of the “errors” were the reference text being wrong. And the Swahili results in a separate 13-language benchmark show what volume alone buys: going from 1 to 400 hours more than halved the best error rate, yet 11.36% is still well above Whisper’s English score from section 3.
| Metric | Value | Source |
|---|---|---|
| Kinyarwanda training data needed for word error rate below 13% | As little as 50 hours | Akera et al., Kinyarwanda and Kikuyu data-scaling study 2025 |
| Kinyarwanda word error rate at 200 hours of training data | Below 10% | Akera et al., Kinyarwanda and Kikuyu data-scaling study 2025 |
| High-error cases caused by noisy reference transcriptions | 38.6% | Akera et al., Kinyarwanda and Kikuyu data-scaling study 2025 |
| African languages benchmarked, with fine-tuning subsets of 1, 5, 10, 20, 50, 100 and 200 hours | 13 languages (Swahili alone also had a 400-hour subset) | Nahabwe et al., Benchmarking ASR Models for African Languages 2025 |
| Best Swahili word error rate with 400 hours of fine-tuning data (XLS-R vs Whisper) | 11.36% vs 16.88% | Nahabwe et al., Benchmarking ASR Models for African Languages 2025 |
Context: the 13 languages in the Nahabwe et al. benchmark are Afrikaans, Akan, Bambara, Bemba, Ewe, Kinyarwanda, Lingala, Luganda, Shona, Swahili, Wolof, Xhosa and Zulu.
5. Crowdsourced and Community Speech Datasets
Open, community-built datasets are where long-tail speech data becomes public. Mozilla Common Voice’s Scripted Speech dataset grew about 31-fold, from 1,368 hours in its first release to 42,593 hours in release 27.0 (42,593 / 1,368 = 31.1), and the repository lists 295 languages (Mozilla, Common Voice cv-dataset repository 2026). The bottleneck has shifted from recording to checking: about 31% of those hours are not validated (29,295 / 42,593 = 68.8% validated).
Regional projects show the same pattern at a larger scale. Project Vaani has gathered over 31,000 hours of Indian speech, yet transcribed only about 6.8% of it (2,122.21 / 31,278.37), and has reached 165 of India’s 773 districts (21.3%). In Africa, two projects reported in 2025 hold more than 10,800 hours between them (9,000+ plus 1,800), against the 5.4 hours of Swahili in Whisper’s speech recognition data.
| Metric | Value | Source |
|---|---|---|
| Common Voice Scripted Speech total hours: release 1 vs release 27.0 (data cut-off September 2026) | 1,368 vs 42,593 hours | Mozilla, Common Voice cv-dataset repository 2026 |
| Common Voice Scripted Speech validated hours, release 27.0 | 29,295 (68.8% of total) | Mozilla, Common Voice cv-dataset repository 2026 |
| Languages listed for Common Voice Scripted Speech: release 19.0 (2024) vs release 27.0 (2026) | 131 vs 295 | Mozilla, Common Voice 19.0 release 2024; cv-dataset repository 2026 |
| Common Voice Spontaneous Speech release 5.0 | 541 hours (302 validated), 80 languages | Mozilla, Common Voice cv-dataset repository 2026 |
| Project Vaani speech collected in India | 31,278.37 hours from 156,208 speakers in 105 languages, 165 of 773 districts | IISc and ARTPARK, Project Vaani dashboard 2026 |
| Project Vaani transcribed speech | 2,122.21 hours (about 6.8% of collected audio) | IISc and ARTPARK, Project Vaani dashboard 2026 |
| African Next Voices recorded speech | 9,000+ hours in 18 languages across 3 countries | University of Pretoria, African Next Voices news 2025 |
| NaijaVoices speech-text data | 1,800 hours from 5,000+ speakers in Igbo, Hausa and Yoruba | NaijaVoices dataset paper, Interspeech 2025 |
Context: Project Vaani’s stated target is more than 150,000 hours of speech from all 773 districts, so the current dashboard represents about a fifth of the planned audio (31,278.37 / 150,000 = 20.9%).
6. Approaches: Self-Supervised Pre-Training, Religious-Text Readings and Few-Shot Extension
The long tail is being reached by three tactics. Self-supervised pre-training (the model first learns from untranscribed audio, then needs only a small labeled set) is the base layer; Google’s Universal Speech Model beat Whisper by 66% relative word error rate on 62 FLEURS languages while using a labeled set about one seventh the size (Google, USM paper 2023). Readings of publicly available religious texts gave the MMS project an average of 32 hours per language, an average that sits inside the 10 to 50 hour band from section 2.
The newest tactic is few-shot extension: give a model a handful of paired audio and text examples when it runs and let it adapt without retraining. The wider picture of shared models that handle many languages at once is covered in our multilingual speech foundation model statistics. Data efficiency also differs by design: with 1 hour of Swahili, MMS, which is built on self-supervised pre-training, beat Whisper by 26.52 points of WER (51.33 - 24.81).
| Metric | Value | Source |
|---|---|---|
| MMS labeled data built from religious-text readings | 44.7K hours across 1,107 languages, about 32 hours per language on average | Meta AI, MMS paper and blog 2023 |
| Character error rate on languages unseen in training: CTC baseline model vs zero-shot model (no retraining) given 10 context examples | 26.33% vs 14.4% | Meta FAIR, Omnilingual ASR paper 2025 |
| USM supervised training data vs Whisper’s, English excluded | 90k hours vs 117k hours, with Whisper beaten by 66% relative WER on 62 FLEURS languages | Google, USM paper 2023 |
| Swahili word error rate after 1 hour of fine-tuning: MMS / W2v-BERT / Whisper / XLS-R | 24.81% / 32.36% / 51.33% / 58.23% | Nahabwe et al., Benchmarking ASR Models for African Languages 2025 |
| Average word error rate improvement after fine-tuning on NaijaVoices: Whisper / MMS / XLSR | 75.86% / 52.06% / 42.33% | NaijaVoices dataset paper, Interspeech 2025 |
Outlier note: the USM paper describes its labeled YouTube set as covering 73 languages in the introduction and 75 in section 3.1. Most recent available data for MMS and USM: 2023.
Summary: Low-Resource Language Speech Recognition Coverage by the Numbers
| Metric | Value | Source |
|---|---|---|
| Languages in use worldwide | 7,170 | Ethnologue, How Many Languages 2026 |
| Languages covered by the broadest open ASR system | 1,600+, over 500 never served before | Meta FAIR, Omnilingual ASR paper 2025 |
| Low-resource languages (under 10 hours) with character error rate below 10 | 195 of 546 (36%) | Meta FAIR, Omnilingual ASR paper 2025 |
| Whisper training audio that was English speech recognition | 65% (438,218 hours) | OpenAI, Whisper paper 2022 |
| Languages spoken across Africa | More than 2,000 | Nahabwe et al., Benchmarking ASR Models for African Languages 2025 |
| Kinyarwanda data needed for word error rate below 13% | As little as 50 hours | Akera et al., data-scaling study 2025 |
| Kinyarwanda word error rate at 200 hours | Below 10% | Akera et al., data-scaling study 2025 |
| Best Swahili word error rate: 1 hour vs 400 hours of fine-tuning data | 24.81% vs 11.36% | Nahabwe et al., Benchmarking ASR Models for African Languages 2025 |
| Common Voice Scripted Speech hours, release 27.0 | 42,593 total, 29,295 validated | Mozilla, Common Voice cv-dataset repository 2026 |
| Common Voice Scripted Speech languages listed | 295 | Mozilla, Common Voice cv-dataset repository 2026 |
| Project Vaani audio collected vs transcribed | 31,278.37 vs 2,122.21 hours | IISc and ARTPARK, Project Vaani dashboard 2026 |
| IndicVoices audio | 7,348 hours, 16,237 speakers, 22 languages | IndicVoices dataset paper 2024 |
| African Next Voices recorded speech | 9,000+ hours, 18 languages | University of Pretoria, African Next Voices news 2025 |
| NaijaVoices speech-text data | 1,800 hours, 3 languages | NaijaVoices dataset paper, Interspeech 2025 |
| USM vs Whisper on 62 FLEURS languages | 66% lower relative WER | Google, USM paper 2023 |
Methodology and Sources
- Ethnologue, How Many Languages Are There in the World? (page read October 2026): living languages, endangerment share, top-20 language concentration.
- OpenAI, Robust Speech Recognition via Large-Scale Weak Supervision (Whisper paper), 2022: training hours per language (Figure 11) and FLEURS word error rates (Table 13).
- Meta, Scaling Speech Technology to 1,000+ Languages (MMS paper), 2023, with the MMS announcement; and Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages, November 10, 2025.
- Google, Universal Speech Model paper, 2023, and FLEURS benchmark paper, 2022.
- Mozilla, Common Voice cv-dataset repository (release 27.0, data cut-off September 2026), the Common Voice 19.0 release post (announced September 18, 2024), and the Scripted Speech 26.0 release notes (June 19, 2026).
- IISc and ARTPARK, Project Vaani live dashboard (read October 2026).
- IndicVoices dataset paper, March 4, 2024.
- Akera et al., How much speech data is necessary for ASR in African languages?, October 8, 2025.
- Nahabwe et al., Benchmarking Automatic Speech Recognition Models for African Languages, November 30, 2025.
- The NaijaVoices Dataset, accepted at Interspeech 2025.
- University of Pretoria, African Next Voices featured in Nature, August 1, 2025.
- Data watch: Language counts differ by source and method (Ethnologue’s 7,170 counts languages in use; ASR papers count ISO 639-3 codes and exclude signed and unwritten languages), so coverage shares are approximations. The Common Voice language count jumped from 131 (release 19.0) to 295 (release 27.0); the repository does not say whether the newer figure counts only languages with released audio, so treat it as languages listed. The 68.8% validated share is computed from the repository’s chart data, while the 26.0 release notes describe about 25% of Scripted Speech audio as still awaiting validation; the two measures are not identical. When release 19.0 was announced, a forum comment noted that only 129 of its 131 languages could be downloaded, so listed and released counts can differ. African Next Voices figures come from a university news page summarizing Nature coverage, and we did not read a dataset paper. Whisper (2022), FLEURS (2022), MMS (2023) and USM (2023) figures are all older than three years and are flagged as most recent available data. The Omnilingual paper’s text (549) and its Table 8 (546) disagree on the number of low-resource languages; we use the table values.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when Mozilla publishes Common Voice Scripted Speech release 28.0.