Mistral: Voxtral Mini TTS

Mistral Voxtral Mini TTS is a text to speech model from Mistral, listed on api14 under the speech category. It takes text as its only input and produces speech as its only output, so it sits in the synthesis half of a speech pipeline rather than the recognition half. If you already route other Mistral models through api14, this one comes from the same provider and follows the same access pattern.

The model accepts text and returns natural sounding audio. It supports multilingual output, and the available voices cover English, British English and French. The English voices are grouped under en_paul, with variants for sad, neutral, happy, frustrated, excited, confident, cheerful and angry. British English offers gb_oliver and gb_jane, each with their own set of emotional variants such as curious, confident, cheerful, angry, sarcasm, confusion, shame, jealousy and frustration. French is served by fr_marie, with sad, neutral, happy, excited, curious and angry variants. The context window is 4096 tokens, and the maximum output is 3276, which is the budget your input text and the generated speech share, so keep source text reasonably short per request.

Those capabilities point to a few practical uses. You can narrate articles, documentation or product copy and pick a voice whose emotional tone matches the material, for example a cheerful voice for onboarding content or a neutral one for technical reading. Because voices exist in three language variants, you can produce localized audio for English, British English and French audiences from the same source text. The emotional range makes the model usable for dialogue in games or e learning, where a single line may need to read as excited, frustrated or sad. Zero shot voice cloning, as described by the provider, is relevant if you need a consistent custom voice across a series of clips.

Calling it through api14 works like any other OpenAI compatible model. Send a request to the chat completions endpoint with the model id "mistralai/voxtral-mini-tts-2603" and your text as the input, and the response contains the generated speech. The model lists no supported parameters, so there are no sampling or generation knobs to tune on the request; the voice selection is the main control you have, and you choose it from the voice list above. Plan your integration around the fixed 4096 token context and 3276 token output ceiling rather than around adjustable settings.

Because the parameter list is empty, this model suits straightforward synthesis jobs where you supply text and a voice and accept the default behavior. If you need fine control over generation settings, or if your input text regularly exceeds the context window, you may want a different model or a chunking step in front of this one. For plain text to speech in the supported languages and voices, it covers the task directly.