Mistral has launched Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK that combines listening, reasoning, tool calling, and speaking in one real-time streaming pipeline. Built for contact centers, service desks, virtual assistants, and hands-free copilots, it is available now through Mistral's EU, US, and global API endpoints. Companies can also deploy the stack in a private cloud or on their own infrastructure.
Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵 pic.twitter.com/YA9HNVfq0u
— Mistral AI (@MistralAI) October 20, 2026
Voxtral Small Transcribe 3 covers more than 50 languages, with diarization, context biasing, and end-of-turn detection. In Mistral's 45-language FLEURS chart, it records forced and inferred error rates of 4.1% and 4.0%, matching Microsoft AI Transcribe 2 and outperforming the listed results for Gemini 3.5 Transcribe, Scribe v2, GPT Transcribe, and Whisper large-v3. The 4-billion-parameter Voxtral Mini Transcribe 3 costs $0.0035 per minute for batch use, while the realtime version costs $0.008 per minute and runs at sub-500 millisecond latency. The transcription models accept key terms and free-form context to raise recall for technical vocabulary without fine-tuning.
Mistral has released new Speech models! Voxtral Small Transcribe 3 takes the #3 spot on the Artificial Analysis Speech to Text leaderboard, and Voxtral 3 TTS enters the top 12 in both Text to Speech arenas@MistralAI has released Voxtral 3: a new Text to Speech model and three… pic.twitter.com/lEqC9AKZIi
— Artificial Analysis (@ArtificialAnlys) October 20, 2026
Voxtral Mini TTS 3.0 delivers batch and realtime speech in more than 35 languages and over 50 dialects for $13 per million characters. It interprets context to adjust emotion, pace, and emphasis, and supports prompt instructions, contextual TTS, a vocabulary library, and voice cloning from more than 30 minutes of audio. For professional deployments, Mistral's Pro Voice adaptation uses 30 minutes of audio to reproduce a speaker's character, emotional range, and speaking style across languages.
The LiveKit-based SDK runs transcription, reasoning, tool use, and speech generation concurrently. A general model such as Mistral Small can reason and call tools while the voice models process the conversation, keeping turns under one second and stopping the agent when a caller cuts in. Agents can verify identities, check bills, book appointments, search the web, and retrieve data from private systems over SIP or WebRTC. Because each stage is separate, teams can change the language model, voice, or language without rebuilding the agent.
Mistral is positioning deployment control alongside model performance. Customers using self-hosted installations can keep audio inside their environment and choose whether conversations are retained, while hosted users can work through Mistral's endpoints. The Studio playgrounds now provide testing for voice agents, transcription, speech generation, and voice cloning before an API or private deployment.
Sources and related context
- Mistral Voxtral 3 speech-to-text documentation: Related context: The guide distinguishes offline transcription from realtime sessions and explains speaker labels, context biasing, and turn detection.
- Mistral Voice Agents SDK documentation: Supports: The documentation supports the article's description of a LiveKit-based voice pipeline with interruptible speech and tool calling.
