Voices
Every speech request names a voice, and the voice decides who the listener hears. Knowing what a voice fixes, and what each request can still change, tells you which kind of voice to use and how to direct it.
What a voice is
A voice is a speaker identity that you reference by its ID in a speech request. It sets how the speaker sounds, their timbre and accent, and it stays the same from one request to the next. That's how your application keeps one recognizable speaker across all its answers.
Everything else comes with each request: the text to speak, how it's delivered, and how specific words are pronounced. To browse the voices and hear them, see the voice library.
Kinds of voices
There are three kinds of voices. They differ in where the voice comes from and how long you can reuse it.
| Kind | Where it comes from | How long you can use it |
|---|---|---|
| Preset voice | Mistral provides it. Voxtral 3 comes with its own voices, and the Voxtral 2 presets are still accepted. | In any request, by its voice ID |
| Cloned voice | A recording that you send with the request. With zero-shot cloning, a short recording is enough and no training is involved. | That request only |
| Saved voice | A cloned voice stored in your workspace | In any request, by its ID, until you delete it |
A clone copies more than the timbre of the speaker. It also picks up the delivery of the recording, such as its emotion and pace, and the way it sounds, such as background noise and the microphone. Record the clip in the style you want to hear. For the recording to provide, see Voice cloning.
Professional voice cloning (PVC) builds a voice from a collection of recordings (30 minutes to 3 hours recommended). The enrolled voice is referenced by its voice ID in synthesis requests. PVC is the recommended workflow for custom voices.
Choose a voice, then direct it
Choose the voice for who should be speaking. The following controls then change how it says each line, without changing the voice:
| Control | What it changes | Scope |
|---|---|---|
| Emotion and pace tags | The emotion and the speaking rate | Part of the text |
| Style instructions | Style instructions are not available at launch. | The whole input |
| Custom pronunciation | How specific words are said, such as brand names | The words you list |
| Expressiveness | Controls how expressive the speech is. Higher values trade voice similarity and stability for stronger delivery. | The whole request |
Emotion in Voxtral 2 and Voxtral 3
Voxtral 2 and Voxtral 3 handle emotion differently. With Voxtral 2, the emotion comes from the voice: each preset speaker has one variant per emotion, such as gb_jane_neutral and gb_jane_sad, derived from its reference audio. With Voxtral 3, you keep one voice and set the emotion inline with tags, which gives finer control. If you're moving from Voxtral 2, see Migrate from Voxtral 2.
Voices and languages
Each preset voice has a language and an accent of its own, listed in the voice library: for example, Jane speaks British English and Marie speaks French. For the languages that speech generation supports, see Languages.