Mistral AI introduces Voxtral 3 for transcription, speech synthesis and voice agents

Mistral Voxtral 3 is a new suite of tools for working with speech, introduced by Mistral AI. It includes speech-to-text models for transcribing spoken words, text-to-speech models for speech synthesis, and an SDK for building voice agents for streamed, real-time conversations.

The service is available through the Mistral API. The company also states that it can be deployed on a customer’s own infrastructure or in a private cloud, which may be relevant for organizations that want to operate audio processing in their own environment.

Mistral Voxtral 3 combines speech, models and tools

The release includes a voice agent SDK built on the LiveKit platform. According to Mistral, it connects voice input, language-model reasoning, tool calling and voice responses into a single workflow. The SDK is intended to support the SIP and WebRTC communication protocols.

This setup is intended, for example, for contact centers, telephone agents and voice assistants. Developers do not have to assemble individual parts of a voice conversation as separate services, but can integrate them through a single SDK.

Language coverage varies by component

Mistral cites broad language support for Voxtral 3, but the exact numbers depend on the model used. Current documentation states:

• 50 languages for batch speech transcription,
• 45 languages and dialects for real-time transcription,
• 35 languages and more than 50 dialects for speech synthesis.

The “50+ languages” label therefore cannot automatically be applied to the entire product suite. It mainly concerns batch transcription; according to the documentation, TTS has separately defined coverage of 35 languages.

Pricing starts with batch transcription

Mistral lists pricing starting at $0.0035 per minute for batch transcription. Real-time transcription costs $0.008 per minute. In its announcement, the company also presents claims about quality, latency and voice-synthesis characteristics, but these parameters have not been independently verified.

For voice agents, practical performance will depend not only on speech-recognition accuracy, but also on the reliability of tool integrations and the system’s behavior when the speaker interrupts. Mistral declares support for real-time conversations, but independent comparisons of the latency of the complete conversational loop are not yet included in the available materials.

What to watch next

Further evaluation of Voxtral 3 will depend on independent benchmarks of transcription accuracy, latency and speech-synthesis quality. Practical experience with tool calling in production deployments and with interrupting agents during calls also remains to be seen.

Verified and updated: October 23, 2026 06:27