Azure’s published text-to-speech table lists 82 distinct language codes, about 1.1% of the 7,170 languages Ethnologue counts in use today (82 / 7,170). Gemini 3.8 Flash TTS documentation claims over 130 languages, and the MMS research project reached 1,107, yet that widest model ships under a non-commercial license. Depth is thinner than breadth: US English has 118 entries in the Azure table against 2 for Amharic. We aggregated data from Ethnologue, Microsoft, Google, Amazon, Apple, Meta, and peer-reviewed speech research, reading every vendor page on October 31, 2026. This roundup covers synthesis only; for the recognition side of the same gap, see our low-resource language speech recognition coverage statistics.
TL;DR
- 7,170 languages are in use worldwide, and 4,153 of them have a developed writing system (Ethnologue 2026).
- 128 of 7,829 language codes reached the top two digital support levels, about 1.6% (Simons, Thomas and White, COLING 2022).
- Over 130 languages are claimed for Gemini 3.8 Flash TTS, the widest commercial claim we read (Google, vendor docs).
- 82 language codes across 154 locales appear in the Azure Speech voice table (Microsoft, vendor docs, our count).
- 25 languages and 42 variants appear in the Amazon Polly table, with 2 variants offering long-form voices (Amazon, vendor docs, our count).
- 1,107 languages have MMS synthesis models, about 15.4% of languages in use (Meta AI, MMS paper 2023).
- 945 of those 1,107 models reached a character error rate of 5 or lower, but Africa trailed at 76% (Meta AI, MMS paper 2023).
- 41 of 90 Aura-2 voices are English (Deepgram, vendor docs, our count).
- 18 of India’s 22 scheduled languages appear in Apple’s accessibility speech list, against 1 in Polly (Apple and Amazon, vendor docs, our tally).
- 1,704 hours from 10,496 speakers now cover all 22 scheduled Indian languages for speech synthesis (AI4Bharat, IndicVoices-R 2024).
1. The Denominator: 7,170 Languages, and Not All of Them Written
Every coverage claim needs a denominator, and synthesis has a harsher one than most people assume. Text-to-speech reads text, so 3,017 languages without a developed writing system sit outside the normal input path before any model is trained.
Population tells a kinder story, and the two views should never be blended. A platform can serve most people while listing under 3% of languages, because over 88% of people speak one of just 200 languages natively, and 200 / 7,170 is 2.8%. None of the vendor pages we read states what share of people its catalog reaches, so reach cannot be compared directly.
| Metric | Value | Source |
|---|---|---|
| Languages in use worldwide | 7,170 | Ethnologue, How Many Languages 2026 |
| Languages with a developed writing system vs without one | 4,153 (57.9%) vs 3,017 (42.1%, from 7,170 - 4,153, our math) | Ethnologue FAQ, 29th edition 2026 |
| People whose native language is one of the 200 largest languages | Over 88% | Ethnologue 200, 2026 |
| Language codes at the top two digital support levels (Vital and Thriving) | 128 of 7,829 (95 + 33), or 1.6% | Simons, Thomas and White, COLING 2022 |
| Language codes at the lowest digital support level (Still, a score of 0) | 3,996 of 7,829, or 51.0% | Simons, Thomas and White, COLING 2022 |
| Language codes at the two middle digital support levels (Emerging, Ascending) | 3,304 and 401 | Simons, Thomas and White, COLING 2022 |
Most recent available data for digital support levels: Simons, Thomas and White, 2022. That study scraped supported-language lists from 143 digital tools, and speech processing is one of its seven tool categories. Totals come from Ethnologue’s language count and its writing system FAQ.
2. Commercial Catalogs Side by Side
These are vendor documentation counts: the primary record of what each platform lists, not independent proof of quality. The gap between the longest and shortest catalog in this table runs from 130+ languages to 9.
The unit matters as much as the number. A locale is a language plus a region, so Spanish for Mexico and Spanish for Spain count as two locales but one language. The Azure overview page says “100+ languages and locales”, which blends the two; the Azure Speech language table itself resolves to the row below.
Longer is also not a superset of shorter. The Gemini TTS language table has no row for Ukrainian, Urdu or Zulu, all three of which Azure lists.
| Metric | Value | Source |
|---|---|---|
| Gemini 3.8 Flash TTS | Over 130 languages claimed; 132 table rows, read October 31, 2026 | Google, Gemini API speech generation docs, updated October 28, 2026 (vendor claim) |
| Eleven v4 / Eleven v3 / Flash v2.5 / Multilingual v2 | 90+ / 70+ / 32 / 29 languages | ElevenLabs, models docs (vendor claim) |
| Azure Speech prebuilt voices | 82 language codes across 154 locales; English, Spanish and Arabic hold 52 of the 154 (33.8%) | Microsoft, language support table, updated October 22, 2026 (vendor docs, our count) |
| Apple VoiceOver, Live Speech, and Read & Speak | 75 language entries on iOS 27 | Apple, iOS feature availability (vendor docs) |
| Google Cloud Text-to-Speech voice table | 53 language codes, 61 language-and-region codes; Chirp 3: HD alone covers 47 languages in 53 locales | Google Cloud, voice list and Chirp 3: HD docs (vendor docs, our count) |
| Sonic 3.6 | 44 languages | Cartesia, TTS models docs (vendor claim) |
| Amazon Polly | 25 languages, 42 variants, 123 voice listings | Amazon, Polly available voices (vendor docs, our count) |
| IBM Text to Speech | 9 base languages, 15 language entries, 48 voices | IBM Cloud docs, updated June 4, 2026 (vendor docs, our count) |
If you are choosing an AI voice generator for a language other than English, the table worth opening is the per-locale list, not the headline number on a product page.
3. Voice Depth: Where the Voices Actually Sit
A language being listed says little about how many voices it gets. US English has 118 entries in the Azure table, while Amharic, Afrikaans and isiZulu have 2 each, a 59-to-1 ratio. For a producer, 2 voices can mean no choice of age, tone or character at all.
The pattern repeats in every catalog size. Polly, Aura-2 and the open Kokoro model put between 28.5% and 51.9% of their voices in English. Part of the difference is structural: table entries can repeat one persona across quality tiers, so these are listing counts, not unique speakers.
| Metric | Value | Source |
|---|---|---|
| Azure voice table entries: US English vs Mandarin vs German vs Hindi | 118 vs 59 vs 23 vs 17 | Microsoft, language support table (vendor docs, our count) |
| Azure voice table entries for Amharic, Afrikaans, Egyptian Arabic and isiZulu | 2 each | Microsoft, language support table (vendor docs, our count) |
| Polly voice listings in English variants | 35 of 123 (28.5%) | Amazon, Polly available voices (vendor docs, our count) |
| Aura-2 voices: English vs Spanish vs French | 41 vs 17 vs 2, of 90 across 7 languages | Deepgram, TTS models docs (vendor docs, our count) |
| Kokoro-82M voices in American and British English | 28 of 54 (51.9%), against 1 for French | Hexgrad, Kokoro-82M voice list (project docs) |
| Google Cloud entries for Hindi (India) vs Icelandic | 46 (30 Chirp 3: HD, 4 Neural2, 6 Standard, 6 WaveNet) vs 1 Standard voice | Google Cloud, voice list (vendor docs, our count) |
| OpenAI built-in voices on the speech endpoint | 13, which the docs call “currently optimized for English” | OpenAI, text-to-speech guide (vendor claim) |
The Cloud voice list shows why tiers matter too: the 30 Chirp 3: HD entries for Hindi come from one shared voice set, and languages outside that set keep older voice types. The Polly count comes from the voice table, where 4 names repeat because they are bilingual, leaving 119 distinct names.
4. Feature Tiers: Listed Does Not Mean Equal
The catalog total is the widest number a vendor can quote, and each premium feature sits on a smaller subset. Polly’s generative engine reaches 25 of 42 variants, and Mandarin, Japanese, Arabic and Russian are among the 17 left out.
Azure shows the same narrowing. Its 154 locales fall to 102 entries for visemes (mouth-shape data used to animate a face) and 42 for cross-lingual transfer (making a custom voice speak a second language). Teams planning a multilingual rollout should check the feature, not the language. Speed and price by model are a separate question, covered in our TTS inference cost statistics.
| Metric | Value | Source |
|---|---|---|
| Azure locales by feature: prebuilt voices / viseme entries / professional custom voice / personal voice / cross-lingual source | 154 / 102 / 68 / 65 / 42 | Microsoft, language support table (vendor docs, our count) |
| Polly variants with at least one generative voice vs at least one long-form voice | 25 of 42 (59.5%) vs 2 of 42 (US English, Spain Spanish) | Amazon, Polly available voices (vendor docs, our count) |
| Polly variants offering only the Standard voice engine | 6 of 42, including Russian, Romanian and Icelandic | Amazon, Polly available voices (vendor docs, our count) |
| Gemini table rows supported by Flash TTS vs Flash-Lite TTS | 132 vs 103; Finnish, Swedish, Thai and Swahili are among the 29 missing from Flash-Lite | Google, Gemini API speech generation docs (vendor docs, our count) |
| Apple Personal Voice languages vs accessibility speech list entries | 3 vs 75 on iOS 27 | Apple, iOS feature availability (vendor docs) |
Apple’s gap is the sharpest: its feature availability page lists Personal Voice only for US English, mainland Mandarin and Mexican Spanish.
5. Open Models: Wide Reach, With Conditions
The open side holds the widest number in this roundup. MMS-TTS reached 1,107 languages, roughly 8.5 times the 130 claimed for Gemini (1,107 / 130), and about 11 times the roughly 100 languages speech technology covered before it (1,107 / 100).
Breadth carried a cost. The MMS paper trained one separate VITS model per language, mostly from single-speaker recordings of religious texts, and its model card lists a non-commercial license. Its most useful result is regional: quality fell furthest in Africa, which the authors attribute “in part” to writing scripts.
Character error rate here means a speech recognizer transcribes the synthetic audio and errors are counted, so lower is better. It measures whether words survive, not whether speech sounds natural; for that side, see our neural TTS prosody and intonation statistics.
| Metric | Value | Source |
|---|---|---|
| MMS-TTS language coverage and license | 1,107 languages (15.4% of 7,170), CC-BY-NC 4.0 | Pratap et al. (Meta AI), MMS paper 2023; Hugging Face model card |
| MMS languages reaching a character error rate of 5 or lower on synthetic speech | 945 of 1,107 (85%) | Pratap et al. (Meta AI), MMS paper 2023 |
| Share of MMS languages reaching that threshold, by region | Africa 76% (277 of 363), Asia 88%, North America 87%, Pacific 90%, South America and Europe 95% | Pratap et al. (Meta AI), MMS paper 2023 |
| MMS character error rate, synthetic vs human reference speech; mean opinion score (1 to 5) on 61 languages | 3.5 ±0.2 vs 2.2 ±0.1; 3.52 for synthetic vs 3.33 for human reference audio outside the training data, a gap the authors attribute to noisy reference recordings | Pratap et al. (Meta AI), MMS paper 2023 |
| SeamlessM4T target languages: speech output vs text output | 36 vs 96 | SEAMLESS Communication Team (Meta), Nature 2025 |
| Rule-based eSpeak NG synthesizer | More than 100 languages and accents | eSpeak NG project README |
| Piper voice list | 37 languages, 42 locale entries; repository archived October 6, 2025 | Rhasspy, Piper VOICES.md |
| Chatterbox Multilingual vs XTTS-v2 | 23+ vs 17 languages | Resemble AI, Chatterbox repository; Coqui, XTTS-v2 model card |
Most recent available data for language-level synthesis quality: Pratap et al., 2023. Newer open models in this table list far fewer languages. The SeamlessM4T paper shows the same asymmetry inside one model: it reads and writes far more languages than it speaks.
6. Africa and India: Where the Gap Gets Specific
Regional tallies expose what global totals hide. Gemini’s table holds 22 languages from Africa, Azure’s holds 5, and Polly and Apple list none beyond South African English, while Sub-Saharan Africa alone is home to over 2,000 distinct languages (Google Research, WAXAL 2026).
Yoruba shows that the gap is not only about data. It appears in the open WAXAL and BibleTTS corpora, yet has no entry in the Azure, Polly, Chirp 3: HD or Gemini tables we read. Public data existing does not mean a platform has shipped a voice.
India tells a better story, because one dataset closed a known hole. Earlier datasets stopped at 14 of the 22 scheduled languages, the IndicVoices-R paper reports, often with 1 or 2 speakers per language. Small communities may need less than expected: Pine and colleagues challenged the assumption that neural synthesis requires large datasets.
| Metric | Value | Source |
|---|---|---|
| Languages from Africa, excluding Arabic varieties: Gemini Flash TTS / Gemini Flash-Lite TTS / Azure / Chirp 3: HD / Polly / Apple | 22 / 13 / 5 / 1 / 0 / 0 | Google, Microsoft, Amazon and Apple catalogs (vendor docs, our tally) |
| WAXAL-TTS languages and recorded hours | 17 languages; over 180 hours on the dataset card, over 235 in the paper, over 565 in the blog post | Google Research, WAXAL dataset card, paper and blog 2026 |
| BibleTTS studio-quality recordings for African synthesis | 10 languages, up to 86 hours per language at 48 kHz | Meyer et al., Interspeech 2022 |
| India’s 22 scheduled languages listed: Apple / Gemini Flash TTS / Azure / Chirp 3: HD / Polly | 18 / 17 / 13 / 10 / 1 | Apple, Google, Microsoft and Amazon catalogs (vendor docs, our tally) |
| Earlier Indian synthesis data: IndicTTS | 284 hours, 27 speakers, 14 languages | AI4Bharat, IndicVoices-R paper 2024 |
| IndicVoices-R speech synthesis corpus | 1,704 hours, 10,496 speakers, 22 languages | AI4Bharat, IndicVoices-R paper 2024 |
| Kanien’kéha speech corpus after a community decision to exclude recordings of speakers who had passed away | 3.46 hours kept of 24 recorded | Pine et al. 2022, as reported by Nigatu et al. 2024 |
| FastSpeech2 model on 1 hour of English vs Tacotron2 on 10 hours | Comparable naturalness | Pine et al., ACL 2022 |
Most recent available data for BibleTTS and for the Pine et al. findings: 2022. The Kanien’kéha figure is carried by Nigatu et al., and it is the most useful number in this table: the limit on that corpus was consent, not technology.
Summary: Text-to-Speech Language and Voice Coverage by the Numbers
| Metric | Value | Source |
|---|---|---|
| Languages in use worldwide | 7,170 | Ethnologue, How Many Languages 2026 |
| People whose native language is one of the 200 largest | Over 88% | Ethnologue 200, 2026 |
| Language codes at the top two digital support levels | 128 of 7,829 | Simons, Thomas and White, COLING 2022 |
| Gemini 3.8 Flash TTS languages | Over 130 claimed | Google, Gemini API docs (vendor claim) |
| Eleven v4 languages | 90+ | ElevenLabs, models docs (vendor claim) |
| Azure language codes and locales | 82 and 154 | Microsoft, language support table (our count) |
| Apple accessibility speech list | 75 entries | Apple, iOS feature availability |
| Amazon Polly languages and variants | 25 and 42 | Amazon, Polly available voices (our count) |
| Azure voice entries: US English vs Amharic | 118 vs 2 | Microsoft, language support table (our count) |
| Aura-2 voices that are English | 41 of 90 | Deepgram, TTS models docs (our count) |
| Polly variants with a generative voice | 25 of 42 | Amazon, Polly available voices (our count) |
| MMS-TTS languages | 1,107 | Pratap et al. (Meta AI), MMS paper 2023 |
| MMS African languages at character error rate of 5 or lower | 277 of 363 (76%) | Pratap et al. (Meta AI), MMS paper 2023 |
| Kokoro-82M voices in American and British English | 28 of 54 | Hexgrad, Kokoro-82M voice list |
| IndicVoices-R corpus | 1,704 hours, 22 languages | AI4Bharat, IndicVoices-R paper 2024 |
| Kanien’kéha speech corpus kept vs recorded | 3.46 vs 24 hours | Pine et al. 2022, as reported by Nigatu et al. 2024 |
Methodology and Sources
All vendor pages were read on October 31, 2026. Figures marked “our count” or “our tally” are our own counts of published tables; they remain vendor claims about what is offered, not independent tests that each voice works well.
- Ethnologue (SIL Global): How Many Languages, Ethnologue 200, unwritten languages FAQ.
- Simons, Thomas and White: Assessing Digital Language Support on a Global Scale, COLING 2022.
- Microsoft: Azure Speech language and voice support, text to speech overview.
- Google: Gemini API speech generation, Cloud voice list, Chirp 3: HD, WAXAL blog, WAXAL paper, WAXAL dataset card.
- Amazon: Polly available voices.
- Apple: iOS feature availability.
- Vendor docs: ElevenLabs models, Cartesia TTS models, Deepgram TTS models, IBM voices, OpenAI text-to-speech guide.
- Meta: MMS paper, MMS-TTS model card, SeamlessM4T in Nature.
- Open projects: Kokoro-82M voices, Piper voices, eSpeak NG, Chatterbox, XTTS-v2.
- Research: BibleTTS, Meyer et al., IndicVoices-R, Pine et al., ACL 2022, Nigatu et al., Zeno’s Paradox of Low-Resource Languages.
- Data watch: Vendor catalogs change without notice, so treat every count as a snapshot. Azure entries repeat personas across quality tiers, and the US English group includes multi-talker entries named for French and Chinese. The Gemini table has 132 rows, but 5 languages appear twice by script, leaving 127 names, and we saw no per-language quality data on the page. Swahili appears in the Chirp 3: HD table but we did not find it in the wider voice list, so the 53 code count may be low by one. Ethnologue’s writing system FAQ quotes a 28th edition figure under a 29th edition citation, and Simons et al. use 7,829 ISO codes rather than 7,170 living languages, so the two denominators are not interchangeable. The three WAXAL sources disagree (27 languages and 565+ hours in the blog, 24 languages and 235+ hours in the paper, 17 synthesis languages and 180+ hours on the card). The SeamlessM4T paper says 36 speech languages in its results but “35 in speech form” in its methods. OpenAI’s guide says both 11 and 13 built-in voices, and it lists its current speech models as shutting down on January 6, 2027. Apple’s list groups three features under one heading. Our regional tallies are language-name matches, not linguistic audits, and the 8.5x and 15.4% comparisons use the 7,170 total and a claim of over 130, so both are approximate.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when Ethnologue publishes its 30th edition.