Deepgram widens Nova-3 and Flux TTS
On 23 October 2026 Deepgram's changelog announced two changes to its speech API: the Nova-3 multilingual recognition model now transcribes 55 languages with one model, where the previous one covered 10, with code-switching and an optional language hint; and the Flux text-to-speech endpoint added 41 languages and 230 voices.
Why it matters
A commercial speech service moves from ten to fifty-five languages in a single recognition model, including speech that switches language mid-conversation, and gains 41 languages in synthesis, among them Ukrainian. The changelog also separates what is finished from what is not: nine of the new synthesis languages ship with text normalisation and 32 are in Early Access without it, and the 55-language model and the language hint are not yet available for self-hosted deployments.
What the changelog says. Entries of 23 October 2026 on developers.deepgram.com. Speech-to-text: model=nova-3&language=multi serves a new multilingual model that transcribes 55 languages with a single model; per-word language tags and code-switching work as before; the optional language_hint biases detection toward the languages the caller expects; the previous 10-language model can be pinned by version. Text-to-speech: Flux TTS (/v2/speak) "now speaks 41 more languages with 230 new voices"; the English voices are unchanged; German, Dutch, French, Hindi, Italian, Japanese, Portuguese, Russian and Spanish ship with text normalisation, the other 32 languages are in Early Access without it, so numbers, dates and abbreviations are read as written; the new voices are on the US, EU and India endpoints and not yet on the Australian one. Japanese numerals formatting is added to the multilingual models; Hindi is not supported there. What the record does not claim. The changelog gives no accuracy figure, no comparison and no quality statement for the new languages, and the 55 recognition languages are listed on a separate overview page, not on this one. It is the vendor's own documentation.