Mistral Launches Voxtral 3 Voice Models and SDK for Real-Time Multilingual Agents
Voxtral 3 delivers state-of-the-art speech transcription in 50+ languages, expressive context-aware text-to-speech in 35+ languages, and a real-time voice agent SDK that…
- Voxtral Small Transcribe 3 leads in multilingual transcription, diarization and context biasing on FLEURS and other benchmarks while adding end-of-turn detection.
- Voxtral Mini TTS 3 produces natural, emotionally adaptive speech across 35+ languages and 50+ dialects at $13 per million characters, with Pro Voice cloning from 30 minutes of audio.
- The voice agent SDK integrates realtime transcription, LLM reasoning, tool use like appointment booking, and speech output for live phone or web calls with sub-second turns.
Real-World Voice Agent Example
In the official demo a French-speaking user named Béatrice provides a medical referral number 4420 1187. The agent transcribes the letter requesting a dermatology opinion, books the next available slot for November 3 at 17:45, confirms the booking, and reminds her to bring the letter and carte Vitale.
This shows the end-to-end capability: accurate transcription across languages, understanding of medical context, tool calling to book an appointment, and natural spoken replies all happening in one continuous conversation.
Model Performance and Features
Voxtral Small Transcribe 3 achieves the lowest word error rates on the FLEURS benchmark across 45 languages, leads diarization tests, and improves keyphrase recall when given context terms. The Mini version offers batch transcription at $0.0035 per minute.
Realtime transcription reaches sub-500 ms latency in over 35 languages with native end-of-turn prediction that combines acoustic and semantic cues to avoid interrupting users.
The TTS model interprets conversational context to adjust emotion, pace and emphasis without extra prompts. It outperforms several competitors on pronunciation accuracy, expressiveness and quality per dollar across multiple languages and dialects.
Deployment and Pricing
Voxtral 3 is available today through Mistral's API on European, US and global endpoints or for self-deployment on your own infrastructure. Organizations control where audio is processed and whether conversations are retained.
Transcription pricing starts at $0.0035 per minute for batch and $0.008 for realtime. Speech generation costs $13 per million characters. Full voice agent API with Mistral Small reasoning is $0.04 per minute. On-premises deployment of the complete stack requires contacting the team.
The October 20, 2026 announcement focuses on accuracy, latency, multilingual support and sovereign control without any claims about content moderation or refusals.
Why it matters for uncensored AI
Our analysis: The source fact that Voxtral 3 supports self-deployment on private infrastructure gives developers and organizations direct control over audio processing and conversation retention policies.
This expands practical options for users who want to keep sensitive voice interactions within their own environment rather than relying on a hosted service. Any reasoning step will still follow the policies of the chosen language model.
