Mistral — Voxtral 3 speech models and a voice-agent SDK (20 Oct, ~18:00 CEST)
Mistral launched Voxtral 3, a new family of speech-to-text and text-to-speech models. It comes with a voice-agent SDK, built on LiveKit, that runs listening, reasoning and tool calling, and speaking as one real-time streaming pipeline. On the transcription side, Voxtral Small Transcribe 3 handles 50+ languages (including all 24 official EU languages) and adds speaker diarization, context biasing for domain terms and end-of-turn detection. Mistral says it beats Whisper, ElevenLabs Scribe v2 and Gemini 3.5 Transcribe on its multilingual and diarization charts. The 4B Voxtral Mini Transcribe 3 costs $0.0035/min in batch, and Voxtral Mini Transcribe Realtime 3 promises sub-500 ms latency across 35+ languages for $0.008/min.
On the speech side, Voxtral Mini TTS 3 (sold as Voxtral Speech 3, batch and realtime on the same API) covers 35+ languages and 50+ dialects. It picks up emotion, pace and emphasis from the conversation, and also takes explicit tags such as [emotion=sad][pace=slow]. It's priced at $13 per 1M characters, which Mistral says undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. A new Pro Voice option clones a speaker's voice across languages from about 30 minutes of audio. The SDK lets a general model such as Mistral Small reason and call tools between listening and speaking. Mistral says turns finish in under a second, and the SDK connects to phone systems over SIP and to browsers and apps over WebRTC. Everything is available today on the Mistral API (EU, US and global endpoints) and in the Studio playgrounds. Mistral also offers self-deployment, but no Voxtral 3 open weights or licence had appeared on Hugging Face by 18:15 CEST. All benchmark claims are Mistral's own.
Why it matters: Europe's main frontier lab now offers a full voice-agent stack, from transcription through reasoning to expressive speech, with EU hosting and on-prem deployment. That pitches it straight at contact centres, against OpenAI's realtime voice models and ElevenLabs.