Speaking of Voxtral | Mistral AI

This post details the release of Voxtral TTS, a 4B-parameter text-to-speech model. It highlights its realistic and emotionally expressive speech generation in 9 languages, low latency, and easy voice adaptation. The model's architecture is described as a transformer-based, autoregressive, flow-matching model built on Ministral 3B, comprising a transformer decoder backbone, a flow-matching acoustic transformer, and a neural audio codec. Performance comparisons with ElevenLabs are presented, showcasing superior naturalness and similar Time-to-First-Audio (TTFA) while supporting emotion-steering. The post also details zero-shot cross-lingual voice adaptation capabilities and its application in cascaded speech-to-speech translation.