Text to Speech
Text to speech turns text into natural speech, with preset or cloned voices, inline emotion and pace tags, and custom pronunciation. This page helps you choose the mode, the model, and the features that fit your application.
Choose your mode
Start from the text you have. It decides the mode, and the page to read next:
| What you have | Mode | Start with |
|---|---|---|
| The complete text, such as a script, an article, or a message to read out | Offline | Speech generation |
| The complete text, and a listener who's waiting to hear it | Offline, with the audio streamed as it's generated | Stream speech |
| Text that's still being written, such as the answer of a model | ||
| A conversation where your application answers by voice | Realtime, run by a voice agent | Voice agents |
For what changes between the two modes, see Offline and realtime audio.
Choose your model
Each mode has its own model. Pick the one for the mode you chose:
| Model | ID | Mode | Use it to |
|---|---|---|---|
| Voxtral TTS 3 | voxtral-mini-tts-3-0 | Offline | Generate speech from complete text, with preset or cloned voices, emotion and pace tags, and custom pronunciation. Stream the audio when the listener is waiting. |
| Voxtral TTS Realtime 3 | voxtral-mini-tts-realtime-3-0 | Realtime |
For the languages and regional dialects that each model supports, see Languages.
Features by mode
Offline requests support every speech feature. Each feature links to the page with its settings:
| Feature | Offline | What it does |
|---|---|---|
| Preset voices | Yes | Pick a ready-made voice by its ID. |
| Zero-shot cloning | Yes | Reproduce a voice from a short recording, and save it to reuse by ID. |
| Emotion and pace tags | Yes: 34 emotions, and the slow and fast paces | Direct the delivery inline, word by word, in the text. |
| Style instructions | No (not available at launch) | Describe the delivery of the whole request in plain words. |
| Custom pronunciation | Yes | Spell out how a name or a term sounds, with phonemes. |
| Expressiveness | A number >= 1.0 that controls how expressive the speech is. The default value is set by the service. | Set how strongly the speech follows the voice, the text, and the tags. |
| Streamed response | Yes | Receive the audio in pieces and start playback before the file is complete. |
| Output formats | mp3, wav, pcm, flac, opus | Get the audio in the format your player or pipeline expects. |
To pick a voice, or to decide between a preset voice and a cloned one, see Voices. To get names and terms right, see Custom vocabulary.
Deprecated models
To move your requests to the new model, see Migrate from Voxtral 2.