Voxtral 3

Voxtral 3 is Mistral’s family of speech models, released on 20 October 2026. It turns recorded or live audio into text and generates spoken replies. A separate voice-agent SDK connects those capabilities with a reasoning model and tools. Mistral’s release announcement offers access through its EU-hosted API or self-deployment.

Choose transcription, speech generation or a conversation

Transcription produces text from audio. Mistral offers Mini and Small models for recorded files, plus a realtime model for speech as it arrives. The speech-to-text documentation explains speaker labels, custom vocabulary and turn detection. Speaker labels identify who spoke; turn detection helps decide when the application should answer.

Speech generation works in the other direction: supplied text becomes audio. Voxtral Mini TTS 3 supports preset or cloned voices, pronunciation controls and changes in emotion or pace. The speech-generation guide distinguishes requests containing complete text from realtime conversations.

The voice-agent SDK combines listening, reasoning, tool calls and speech. For example, an application can ask a connected booking system for available times, discuss the options with the caller and request confirmation before making a booking.

Language coverage depends on the task

Mistral reports transcription support for more than 50 languages. Speech generation supports more than 35 languages and 50 dialects. The realtime transcription model has its own supported-language list. Those figures describe different components; check the model used by the application.

The release’s quality and latency comparisons come from Mistral’s evaluation setup. Test recordings, accents, interruptions and pronunciation on the intended workload before selecting a model. The audio directory covers other speech and music models.