Voxtral TTS

What it is

Voxtral TTS is the first text-to-speech model released by Mistral AI, on 23 March 2026 ^[1]. It turns written text into spoken audio, and it can also copy a voice: given a short recording of a speaker, it reads new text in that voice. The model has about 4 billion parameters ^[1].

Text-to-speech is the step that lets a language model's output be listened to rather than read, which is what narration, accessibility tools and voice interfaces are built on.

How it is built

Mistral built the model on Ministral 3B. The system has three parts: a 3.4B Transformer decoder, a 390M flow-matching acoustic transformer, and a 300M neural audio codec running at 12.5 Hz ^[1]. Mistral describes the result as a transformer-based, autoregressive, flow-matching model ^[1].

The three names describe standard roles in speech generation. The decoder works on the text, the acoustic transformer produces the representation of the sound, and the codec is the part that converts between audio and the discrete units the model predicts. Because the codec runs at 12.5 Hz, one second of speech becomes 12.5 units rather than thousands of individual samples, which keeps the sequence the model has to generate short.

"Autoregressive" means the model produces units one at a time, each conditioned on the ones already produced — the same pattern a language model uses for text. "Flow matching" is a way of generating continuous signals: the model learns a transformation that carries a simple starting distribution, essentially noise, towards the target, and the acoustic transformer uses it to turn the decoder's output into audio features.

Languages and voice cloning

It supports nine languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic ^[1]. Voice cloning needs a reference clip as short as three seconds, and the model can speak across languages zero-shot ^[1]. Zero-shot here means no additional training on the target voice or language is involved; the reference clip is the whole input. A three-second threshold means a casually recorded sample is enough, which is what makes one voice usable across several languages in a dubbing workflow.

Speed and length

For a typical request the model latency is 70 ms, and a single generation can produce up to two minutes of audio ^[1]. Latency is the time the model itself takes to answer; two minutes is long enough for a full stretch of narration rather than a single sentence.

How to get it

The weights are open under CC BY-NC 4.0 ^[1]. That licence allows reuse with attribution but excludes commercial use. Through the API the price is $0.016 per 1,000 characters ^[1]. Several reference voices are hosted on Hugging Face under the same licence ^[1].

Evaluations

In Mistral's human evaluation, listeners preferred it to ElevenLabs Flash v2.5 on naturalness ^[1]. Mistral also said Voxtral TTS achieves quality at parity with ElevenLabs v3 ^[1]. Both statements come from the company's own testing rather than from an independent benchmark.

Where it sits

Voxtral TTS was succeeded by Voxtral TTS 3, which went into public preview on 20 October 2026 ^[3]. Voxtral TTS itself is listed as deprecated on 31 October 2026 ^[2][3], with retirement on 30 April 2027 ^[2].