Mistral has unveiled Voxtral 3, a new family of speech models and a voice agent SDK designed to handle real-time conversations across more than 50 languages. The French AI company says the suite combines transcription, reasoning, speech synthesis and tool calling into a single streaming pipeline.
The launch targets contact centers, virtual assistants and hands-free copilots. Voxtral 3 is available today through Mistral's EU-hosted API or for self-deployment on customer infrastructure, keeping audio processing inside controlled environments.
Voxtral 3 speech models and SDK capabilities
The Voxtral 3 family includes two speech models. Voxtral Small Transcribe 3 handles multilingual transcription with speaker diarization, context-biasing for technical terms and end-of-turn detection so agents respond without delay. According to Mistral, it outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe in English, French and Spanish, with word error rates as low as 4.0%.
For speech output, Voxtral Mini TTS 3 generates natural voices that adapt emotion, pace and emphasis to conversational context, covering 35+ languages and 50+ dialects. A 4B-parameter Mini variant of the transcription model transcribes at $0.0035 per minute.
The voice agent SDK ties listening, reasoning and action together. Agents can retrieve information, interact with business systems and complete tasks during live conversations over phone or web. For professional use cases, Pro Voice adaptation reproduces a speaker's distinctive character and emotional range across languages using just 30 minutes of audio.
Multilingual transcription performance
Voxtral Transcribe 3 supports 50 languages, including 24 official EU languages and several low-resource Asian and Indic languages. Both models accept keyphrases and free-form text to bias transcription, improving recall on domain-specific vocabulary without fine-tuning.
