Mistral

Mistral Voxtral generates speech from preset voices in about 35 languages, from voices you have cloned in your Mistral workspace, or from a short recording sent with the request. The same Mistral API key also serves speech recognition. Voxtral tuning is passed in synthesizer.options.

{
  "verb": "say",
  "text": "[emotion=happy] Your order shipped this morning and should arrive on Thursday.",
  "synthesizer": {
    "vendor": "mistral",
    "voice": "voice_478",
    "options": {
      "expressiveness": 3
    }
  }
}

You can also set these in the portal under Extra Options on the speech credential, in which case they apply to every request that uses that credential. Settings supplied on the verb take precedence.

Voices and languages

Each voice speaks its own language, and Voxtral has no separate language setting, so jambonz does not send synthesizer.language to Mistral. Pick a voice for the language you want. Voxtral 3 preset voices have ids such as voice_478 (Sheila, US English) or voice_592 (Marie, French). The Voxtral 2 presets (en_paul_neutral, gb_jane_neutral, …) are still accepted. Voices you have created in your Mistral workspace are used by their id.

The default model is voxtral-mini-tts-3-0. Set a different one with synthesizer.options.model_id, or with the model on the speech credential.

Emotion and pace tags

Put tags inline in the text to change how a passage is spoken, such as [emotion=happy] or [pace=slow], or both together: [emotion=calm][pace=slow]. Pace takes very-slow, slow, moderate, fast or very-fast. Voxtral does not speak a tag it does not recognize, so a misspelled tag is silently ignored. Text is synthesized a sentence at a time, and jambonz carries a reply’s tags into each later sentence until one sets its own, so a tag at the start of a streamed reply shapes the whole reply.

Options

expressiveness

number

How much emotional range the voice uses, 1 – 5.

ref_audio

string

A short recording, base64-encoded, to clone a voice for this request only. Used instead of voice.

phoneme_hints

array

Custom pronunciations, passed to Mistral unchanged as phoneme_hints.

conversation_history

array

Earlier turns of the conversation, oldest first, so the voice matches the tone of the exchange. Turns alternate user and assistant and end with a user turn. Requires a Voxtral 3 preset voice and cannot be combined with ref_audio.

prompt_cache_key

string

A key that lets Mistral reuse work across requests that share it.

Every option is optional. An option you don’t set is not sent, so Mistral’s own default applies.

Endpoint override

To point at a non-default endpoint, such as a proxy, set api_uri in the credential’s Extra Options; it applies to both TTS and STT. Leave it out to use api.mistral.ai.

Streaming

Mistral supports TTS streaming, so it can be driven token by token from an LLM using tts:tokens, for example from the agent verb. Voxtral has no connection that accepts text as it arrives, so jambonz collects the tokens into whole sentences and sends each sentence to Mistral as a separate streaming request. The first audio therefore starts once the first sentence is complete. Splitting at sentence boundaries keeps the intonation natural.

On barge-in, jambonz cancels the request in progress. Voxtral does not return word-level timestamps, so jambonz cannot trim the assistant’s turn to the words the caller actually heard.