Deepgram adds 41 languages and 230 voices to Flux TTS

Deepgram announced additional languages and voices for Flux text-to-speech.

**Dek:** The October 23 update keeps the same batch and streaming speech endpoints. Nine of the new languages include text normalization, while the other 32 arrive in Early Access with a different input requirement.

Deepgram added 41 languages and 230 voices to Flux text-to-speech on October 23. The company says developers can select the new voices using the existing model naming format and the same batch and streaming speech routes. Existing English voices are unchanged, according to the release notes. For teams expanding a voice product, the material change is a wider set of languages within an existing integration. How those languages read written dates and numbers depends on their rollout stage. [Deepgram October 23 changelog](https://developers.deepgram.com/changelog/2026/10/23)

Voice selection and text preparation are different tasks

Deepgram lists German, Dutch, French, Hindi, Italian, Japanese, Portuguese, Russian and Spanish as shipping with text normalization. The other 32 additions are Early Access without that feature. In those cases, Deepgram says numbers, dates and abbreviations are read as written and should be written out in words. A localized service that already stores a numeric value in its database therefore may need to format that value for speech before sending it to the synthesizer. [Deepgram October 23 changelog](https://developers.deepgram.com/changelog/2026/10/23)

Consider an appointment reminder containing a time, a telephone number and a clinic abbreviation. A human reader may know to say the phone number as digits and expand the abbreviation, while the synthesizer only receives the literal characters in the message. The application’s text preparation can supply the intended spoken form. That needs to follow the language and regional conventions of the service. A global substitution rule may read a currency amount correctly in one locale and poorly in another.

What a rollout should measure

Select candidate voices from the current catalogue and evaluate them against the prompts the product actually uses. A short set can cover ordinary statements, urgent instructions, dates, abbreviations, currencies and proper names. Where the product supports both streaming and batch requests, compare audio format, time to first usable audio and how the application handles a failure. The provider’s voice count describes the choices available. It does not tell an operator which voice is clearest for a particular audience. [Deepgram October 23 changelog and catalogue links](https://developers.deepgram.com/changelog/2026/10/23)

One operational benefit of keeping the existing routes is that a team can add language support without immediately replacing the transport layer. The split between normalized and Early Access inputs is still a contract difference. Teams can capture it in their locale configuration, so a new voice becomes a reviewed deployment choice rather than a change to every language at once.