Mistral Launches Voxtral 3 Voice Agent SDK
Original: Voxtral 3
Why This Matters
Unified STT+TTS+tool-calling in one pipeline lowers the bar for production voice AI.
Mistral released Voxtral 3, a voice agent SDK and speech model family supporting 50+ languages, combining real-time STT, TTS, and tool-calling in a single streaming pipeline. Available via EU-hosted API or self-deployment.
Mistral's Voxtral 3 is a new family of speech models paired with a voice agent SDK designed to handle end-to-end voice workflows in real time. The release includes two core models: Voxtral Small Transcribe 3 for speech-to-text, and Voxtral Mini TTS 3 for text-to-speech.
The transcription model covers 50+ languages and adds speaker diarization, context-biasing via keyphrases, and end-of-turn detection—so agents don't awkwardly step on users mid-sentence. The TTS model generates expressive speech that adjusts tone and emotion based on conversational context.
The SDK chains listening, reasoning, tool-calling, and speaking into a single streaming pipeline. A demo scenario shows the system reading a French medical referral letter, booking an appointment, and responding in English—all in one uninterrupted flow.
Mistral is positioning the stack for contact centers, service desks, virtual assistants, and hands-free copilots. Consistent with the company's broader positioning, Voxtral 3 is available through Mistral's EU-hosted API or for self-deployment on private infrastructure, giving enterprise customers control over audio data and retention policies.