Neural ExpeditionSubscribe
TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #95

Turns text into lifelike speech and picks up a new voice from a short sample

Mistral rarely opens up its audio stack, and this is its first open text-to-speech release. If you have wanted one consistent voice across languages, this is the closest an open model has come.

MODELVoxtral-4B-TTS-2603
PUBLISHEDAugust 24, 2026
READ TIME2 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

You give it text and it speaks the line back in a natural, expressive voice. It ships with 20 preset voices, and it can pick up a new voice from a short reference: give it about ten seconds of someone speaking and it reads your text in that voice, in any of its nine languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, and Hindi.

The practical difference is that it is built for live conversation, not just narration. It streams audio as it generates, with the first sound arriving in a fraction of a second, which is what lets a voice agent answer a caller without an awkward pause.

Under the hood it is a 4 billion parameter model built on Mistral's small Ministral language model, and the weights run on a single graphics card with 16 GB of memory. Output is clean 24 kHz audio.

02How to try it

Open the voxtral-tts-demo Space on Hugging Face and type the same two sentences into a few different voices and languages. Listen for whether the voice keeps its character when you switch from English to French or Arabic; that consistency across languages is the part most speech tools get wrong. For your own voice samples or heavier use, run the weights locally with the vLLM server; the setup commands are on the model page and a single 16 GB GPU is enough.

03Caveat

The license is the real limit: the weights and reference voices are released for non-commercial use only, so you can test, research, and prototype freely, but shipping it inside a commercial product requires Mistral's paid API instead. And as with any voice cloning tool, only clone voices you have permission to use.

04What you can do with it

  • Give a multilingual product video one narrator voice across every language version.
  • Build a phone or support agent that answers with the same voice every time.
  • Draft audiobook or podcast narration before booking any studio time.
  • Have your app read notifications or articles aloud in the user's own language.
  • Prototype dubbing: your voice, speaking a language you do not.

Try the demo

View model page

Read this issue on neuralexpedition.com →