Voxtral 3 connects multilingual speech with live business workflows

Mistral has released a voice-agent SDK and new speech models designed to listen, reason, call tools and respond in one streaming interaction.

Mistral's Voxtral 3 release brings speech recognition, synthesis and tool use into a single voice-agent pipeline. Announced on 20 October, the family is available through Mistral's EU-hosted API or self-deployment, giving organisations a choice of operating environment for audio processing.

The release includes Voxtral Small Transcribe 3, with transcription across more than 50 languages, speaker diarisation, context biasing and end-of-turn detection. Voxtral Mini TTS 3 adds expressive speech in more than 35 languages and 50 dialects. The voice-agent SDK connects listening, reasoning, tool calling and speaking in real time.

From recognising speech to completing a request

A contact-centre interaction contains several distinct steps. The system needs to hear a name or reference number, understand a correction, retrieve the relevant information and decide which action the customer is requesting. It then needs to communicate what happened in a form the caller can follow.

Voxtral's SDK is designed to join those steps. A connected tool can look up an appointment or retrieve an account record while the conversation continues. Context biasing supplies important vocabulary, such as product names or domain-specific terms, to the transcription stage. Diarisation separates speakers, while end-of-turn detection helps determine when a reply should begin.

Deployment follows the service workflow

For a service team, the starting point is a defined task and the systems needed to complete it. An appointment workflow needs scheduling access; a support workflow needs the relevant account and case tools. The voice layer then becomes an interface to those existing operations.

The choice between a hosted API and self-deployment also sits within that service design. It affects how audio enters the environment, where processing occurs and how the organisation manages conversation retention.

The release brings the components closer together. Instead of assembling transcription, a reasoning loop, tool execution and speech output as four separate experiences, developers can build around one continuous interaction and concentrate on the service the caller is trying to complete.

Sources

- [Voxtral 3](https://mistral.ai/news/voxtral-3/) — Mistral AI · 2026-10-20 - [Speech to Text documentation](https://docs.mistral.ai/studio/audio/speech-to-text/overview) — Mistral AI