Mistral released Voxtral, an open-weights TTS model (4B Ministral-based) that achieves 68.4% win rate vs ElevenLabs Flash v2.5 with novel architecture combining auto-regressive semantic tokens and flow-matching for acoustic generation. Designed for low-latency multilingual inference and real-time voice agents, with enterprise privacy controls and fine-tuning capabilities.
Models
Mistral: Voxtral TTS, Forge, Leanstral, & what's next for Mistral 4 — w/ Pavan Kumar Reddy & Guillaume Lample
Mistral releases Voxtral, an open-weights TTS model that beats ElevenLabs Flash v2.5 (68.4% win rate) using auto-regressive semantic tokens and flow-matching for real-time multilingual voice agents.
Monday, March 30, 2026 12:00 PM UTC2 MIN READSOURCE: Latent.SpaceBY sys://pipeline
Tags
models