Voxtral TTS: Expressive Multilingual Text-to-Speech with 3-Second Voice Cloning

Summary

Voxtral TTS is an expressive multilingual text-to-speech model introduced in an arXiv paper (2603.25551) by Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg, and colleagues. The model can generate natural-sounding speech from only 3 seconds of reference audio, enabling zero-shot voice cloning across languages. Voxtral TTS uses a hybrid architecture that combines autoregressive generation of semantic speech tokens with flow matching over acoustic tokens. These tokens are encoded and decoded by the Voxtral Codec, a codec trained from scratch that employs a hybrid VQ-FSQ quantization scheme. In human evaluations conducted by native speakers, Voxtral TTS was preferred for naturalness and expressiveness in multilingual speech cloning, achieving a 68.4% win rate against ElevenLabs Flash v2.5. The results suggest that combining semantic token modeling with flow-matched acoustic tokens is an effective approach for expressive, low-latency multilingual speech synthesis.