Voxtral TTS

Voxtral TTS is Mistral AI’s current text-to-speech model for low-latency speech generation and zero-shot voice cloning.

What this model is

The standout feature is voice cloning without requiring a transcript of the reference audio. A developer can supply an authorized voice prompt and generate new speech in that voice, reducing setup compared with systems that need paired audio/text data. Mistral also emphasizes streaming performance with roughly 90ms time-to-first-audio, making Voxtral relevant to interactive assistants as well as offline narration.

Technical capabilities and model behavior

The hosted model ID is voxtral-mini-tts-2603. It converts text to audio, supports nine languages and includes voice-cloning functionality through the speech endpoint. Unlike a general language model, it does not reason, call tools or generate structured text. It is best understood as the audio-rendering layer in a larger system: a language model decides what to say, while Voxtral determines how that text is spoken.

How it works in real applications

A voice assistant can generate the response with Mistral Medium or another LLM and stream the final text into Voxtral TTS. An accessibility product can create natural spoken versions of documents. A localization workflow can generate multilingual narration, while authorized creators can clone a speaker’s voice for approved content. Separating language reasoning from speech generation also allows teams to swap the TTS model without rewriting the agent logic.

Current status and availability

Voxtral TTS became GA on March 23, 2026. Mistral provides hosted API access and released weights. However, the weights use the CC BY-NC 4.0 license, which is non-commercial. That distinction is important: downloadable weights do not automatically mean unrestricted open-source commercial deployment. Commercial users should review the hosted API terms or seek appropriate licensing.

Pricing and deployment considerations

Hosted pricing is $16 per million input characters rather than per text token. Character-based billing makes narration costs easy to estimate from script length. For interactive use, total system cost also includes the upstream language model and any transcription/agent services. Streaming can reduce perceived latency even when full audio generation continues in the background.

Who should choose this model?

Choose Voxtral TTS when you need multilingual speech, fast first audio and authorized voice cloning. A standard no-cloning TTS system may be easier to govern if identity imitation is unnecessary. Teams requiring fully permissive commercial self-hosting should also examine the license carefully rather than assuming the downloadable checkpoint can be used without restrictions.

Important limitations and trade-offs

Voice cloning creates misuse risks that are different from ordinary text generation. Applications should obtain clear speaker consent, communicate when audio is synthetic and protect reference recordings. Speech quality can vary by language, speaking style and unusual names. Production systems should also test pronunciation, latency, audio artifacts and fallback behavior rather than evaluating only a short demo.