Mistral AI has launched Voxtral 3, a new family of speech models paired with a real-time voice agent SDK, positioning the French lab as a one-stop shop for conversational voice AI.
The release includes three transcription models and one text-to-speech model. Voxtral Small Transcribe 3 is the flagship speech-to-text engine, covering more than 50 languages with speaker diarization, context biasing for domain terms, and end-of-turn detection. A lighter 4B-parameter Mini Transcribe 3 handles cheaper batch jobs, while a real-time variant promises sub-500ms latency for live conversations. On the output side, Voxtral Speech 3 generates speech in over 35 languages and 50 dialects, adapting emotion, pacing and emphasis — and a Pro Voice feature can clone a speaker’s voice and style from 30 minutes of audio.
The voice agent SDK, built on LiveKit, ties listening, reasoning, tool calling and speaking into a single streaming pipeline, with each conversational turn completing in under a second. Agents can call external tools mid-conversation — for example, to book appointments or check invoices — over SIP telephony or WebRTC.
Pricing starts at $0.0035 per minute for transcription and $13 per million characters for speech. The models run on Mistral’s EU-hosted API or can be self-deployed, a pitch aimed squarely at companies that want control over where voice data is processed and retained.
The launch puts Mistral in direct competition with established speech APIs such as Deepgram, AssemblyAI and Speechmatics — and Mistral claims its transcription beats Whisper and Gemini on multilingual benchmarks.
Businesses exploring voice automation can compare options in our Best AI Tools for Small Business in 2026 roundup.
Source: The Decoder
