Mistral AI launched Voxtral 3, a new speech model suite paired with a real-time voice agent SDK built on LiveKit. Available in public preview on the Mistral API, the release includes three transcription models and one text-to-speech model. The streaming pipeline operates with under one second of latency per turn and handles automated caller interruptions across WebRTC and SIP phone connections.
The models support diarization and context biasing across 50 languages. Mistral reported that Voxtral Speech 3 costs $13 per 1 million characters and features explicit controls for pacing and 34 emotion tags. The models are hosted across Mistral’s European, US, and global endpoints with zero-data-retention options.
Why it matters
Provides voice developers a modular, low-latency stack with explicit emotion controls to deploy real-time conversational agents.
Gives European enterprises a local, privacy-compliant voice AI alternative with support for 24 official EU languages.
Source: marktechpost.com



