TTS language counts in October 2026: Voxtral 9, MAI 23, ElevenLabs 90+

How many languages each TTS vendor states: Mistral Voxtral, Microsoft MAI-Voice-2.1, OpenAI and ElevenLabs, from vendor pages, plus Sume's language field.

5 min readSume
All posts

The vendors count languages differently, so the numbers do not line up like for like. Mistral's Voxtral TTS lists 9, Microsoft's MAI-Voice-2.1 lists 23 languages (with several English locales), OpenAI says its TTS supports 50+ languages with voices optimized for English, and ElevenLabs lists 90+ for Eleven v4 and 32 for Flash v2.5. All of these come from vendor pages read on 2026-10-04. Sume's TTS takes a language code and does not publish a count in the docs read here.

The table

Counts are what each vendor states about the model, not what we tested.

Stated TTS language coverage by vendor page (read 2026-10-04)
ModelStated coverageSource page
Mistral Voxtral TTS9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, ArabicMistral launch post
Microsoft MAI-Voice-2.1 and Flash23 languages, including several English localesMicrosoft model page
OpenAI TTS (gpt-4o-mini-tts, tts-1, tts-1-hd)50+ languages; voices currently optimized for EnglishOpenAI TTS guide
ElevenLabs Eleven v490+ languagesElevenLabs models overview
ElevenLabs Flash v2.532 languagesElevenLabs models overview
ElevenLabs Multilingual v229 languagesElevenLabs models overview

Why a bigger number is not automatically better

A language count tells you a model will attempt the text. It does not tell you the stress, numbers, names or accent will be right. OpenAI's own wording, voices optimized for English, is a useful reminder: support and quality are separate claims. Microsoft's page lists the languages by name, which makes its claim easy to check against your markets. Mistral lists fewer languages but includes Hindi and Arabic, which matter for some markets and are missing from several shorter lists.

Whichever you pick, test with your own copy. A minimum test is three lines per language: a sentence with a product name, a sentence with a price, and a question.

Where Sume fits

Sume's POST /v1/tts-1.0/generate accepts a language code (BCP-47 or ISO-639) and asks you to set it on every non-English transcript. TTS 1.0 has no engine picker; the TTS router lists explicit catalog models. The router's enum in the OpenAPI file names Sonic models, so check the live catalog rather than assuming a vendor model from this table is on it. Jobs come back as hosted artifacts that the rest of the platform reads, which is covered in the models overview.

If a market is not on your chosen vendor's list, the honest answer is to run a test synthesis, listen, and pick another model for that market. A mixed approach is normal: one vendor for the long tail of languages, another for the three markets that earn the revenue.

How to use the table

Start from the markets, not from the vendor.

  • List the languages you will ship in the next two quarters.
  • Strike every model whose stated list misses one.
  • Audition the rest on your copy, using the plan in one voice in 23 languages.
  • Record the winner per language and re-check when the vendor page changes.

Sources

Related posts

More in Comparisons

All Comparisons posts

Written by Sume