Mistral: Voxtral TTS 3

Mistral Voxtral TTS 3 is a text-to-speech model from Mistral. It sits in the speech category and is identified on api14 as mistralai/voxtral-mini-tts-3-0. It is produced by Mistral, the same provider behind the rest of the Voxtral family, and its job is narrow and specific: turning written text into spoken audio rather than generating or editing text.

The model takes text as its only input and returns speech as its only output. It has a context window of 4096 tokens, and the maximum output is 3276 tokens, which bounds how much text you can send in a single request. Voxtral TTS 3 ships with a large library of preset voices. These include named English voices such as en_paul_neutral, en_paul_happy, en_paul_sad, en_paul_angry and en_paul_excited, British voices such as gb_oliver_neutral, gb_jane_sarcasm and gb_jane_confused, French voices such as fr_marie_neutral and fr_marie_angry, alongside many numbered voices like voice_3150, voice_478 and voice_3132. According to the provider, the model also supports zero-shot voice cloning and accepts emotion and pace tags inside the text, which lets you steer delivery without changing the underlying script.

The practical uses follow directly from those traits. With emotion and pace tags plus a wide set of voices, the model fits narration for videos and courses where the delivery needs to shift between calm, excited or frustrated lines. The British and French voice sets make it usable for localized content without a separate model per locale. Zero-shot voice cloning is useful when you want a consistent brand or character voice across many clips, and the preset library covers cases where you just need a natural voice quickly. Because output is audio only, it works well as the last stage of a pipeline: generate or edit your script with a text model, then pass the final text here for speech.

Calling it through api14 works the same way as any other model on the gateway, since the API is OpenAI-compatible. You send a request to the speech endpoint with the model id mistralai/voxtral-mini-tts-3-0, pass your text as the input, and choose a voice from the list above. Note that the model reports no supported parameters of its own, so you should not expect the gateway to forward arbitrary tuning options such as temperature or top_p; the voice selection, the text and any emotion or pace tags you embed are how you control the result. Keep requests within the 4096 token context and the 3276 token output ceiling.

If your task is generating or rewriting text rather than speaking it, a text model is the better fit, since Voxtral TTS 3 accepts only text and returns only speech. It also is not a transcription model, so audio input is out of scope. For long documents you will need to split the script into chunks that respect the context limit and stitch the resulting audio together.