MistralAI introduces Voxtral 3: Advanced models for voice agents
MistralAI introduced Voxtral 3—a comprehensive suite of cutting-edge speech processing models covering both transcription and speech synthesis. These tools are designed for creating highly functional voice agents.
The Voxtral Small Transcribe 3 model provides advanced multilingual transcription, including diarization (separating speech from different speakers) and contextual shifting, allowing agents to respond without delay.
For speech synthesis, Voxtral Mini TTS 3 is presented, achieving high expressiveness in over 35 languages and 50 dialects. It is capable not only of mimicking emotion, tempo, and accent but also of following explicit instructions given in the form of tags (e.g., [emotion=sad]).
Furthermore, MistralAI introduces Pro Voice, professional voice adaptation, allowing the user's voice to be recreated with high accuracy by recording as little as 30 minutes of audio. All of this is combined into an SDK for voice agents, creating a unified pipeline: transcription -> reasoning (tool calling) -> voice response.
Why it matters
- —Advanced integration: Voxtral 3 creates a complete pipeline from speech recognition to response, mimicking natural dialogue.
- —High emotional expressiveness: The new TTS model can account for emotions, tempo, and accent, which is critically important for realistic voice assistants.
- —Professional voice adaptation: The ability to clone voices with high accuracy opens up new possibilities for enterprise use.
Key facts
- Voxtral Small Transcribe 3 supports multilingual transcription and diarization.
- Voxtral Mini TTS 3 works on 35+ languages and 50+ dialects, mimicking emotions and tempo.
- The Pro Voice function is implemented for professional voice adaptation using recordings up to 30 minutes.
- An SDK for voice agents has been released, combining transcription, reasoning logic, and speech synthesis.
The full text is in the original source. Here we provide a brief summary and key facts.