Supported languages

Overview

Speech is generated in ten languages, on every speech model.

Transcription covers more languages than synthesis, so check both directions before committing to a language list: a product can transcribe a user in a language it has no voice to answer in. Speech-to-text languages has the transcription side.

Voices and languages

A preset voice has a native accent, and that is not a restriction on what it can read. A voice listed as Chinese reads English; it sounds like a Chinese speaker reading English.

Use a preset when the accent does not have to match the output language, as in a global assistant. Match the accent to the language for a localized experience. Accent and a playable sample for each voice are in the voice library.

To get a specific accent in a specific language, use voice design: describe the accent you want and the model synthesizes a voice to match.

Setting the language

language selects the output language. Most requests do not need it, since the language is inferred from input.

Set it when the text is ambiguous: a short string, a proper noun that reads differently in two languages, or a number and date format that could belong to either.