Mistral just open sourced a TTS model!

(3 second voice cloning, beats ElevenLabs!!!)

Voxtral TTS is a 4B parameter multilingual text-to-speech model. It clones any voice from just 3 seconds of reference audio using a hybrid architecture that combines autoregressive decoding for semantics with flow-matching for acoustics.

- 68.4% win rate over ElevenLabs Flash v2.5
- 9 language support with benchmarks
- Outperforms ElevenLabs on speaker similarity across all languages
- Sub-second latency, 32 concurrent streams on a single H200
- DPO adapted for hybrid discrete-continuous generation

Link to the paper and model weights :
https://huggingface.co/papers/2603.25551