Mistral launches Voxtral 3 speech models and a toolkit for voice agents

Mistral introduced Voxtral 3 on October 20, pairing a new speech-model family with a toolkit for building voice agents that can carry out actions during a conversation. The launch brings transcription and speech generation into a wider workflow that also uses a language model and connected tools. Mistral is targeting applications including customer support, service desks and virtual assistants, where understanding a spoken request is only the beginning of the task.

The voice-agent SDK manages audio capture, transcription, model responses and spoken playback. It allows a caller to interrupt an answer and lets an agent retrieve documents, search the web or invoke business functions while the conversation continues. Its transport builds on LiveKit, with Python and TypeScript bindings. The design keeps the application’s knowledge and permitted actions configurable, so the same conversational pipeline can support different services.

The transcription family separates recorded audio from live streams. Voxtral Mini Transcribe 3 and the larger Small Transcribe 3 handle files, while Mini Transcribe Realtime 3 returns text as audio arrives. Mistral’s documentation describes speaker labeling and vocabulary biasing, which supplies names or specialist terminology to improve recognition. For recordings, these features can help preserve who said what without treating every unfamiliar word as a problem requiring a newly trained model.

The distinction between modes also carries concrete limits. Offline transcription offers segment timestamps rather than word-level timestamps. Realtime speaker labeling and end-of-turn detection are each supported, but the documentation says they cannot be enabled together in the same session. End-of-turn detection helps an agent decide when to answer; diarization distinguishes speakers. An application that needs both must account for that separation rather than assuming the launch combines every capability in one connection.

On the output side, Voxtral TTS 3 generates speech from complete text, with a realtime variant for conversational use. The current documentation lists preset voices, cloning, emotion and pace tags, and pronunciation controls. Audio can arrive in pieces before the full file is ready, allowing playback to begin earlier. These controls give developers a way to handle names, delivery and voice identity separately instead of relying entirely on the language model’s written answer.

Some speech features remain more limited than the broad product pitch might suggest. The text-to-speech guide marks free-form style instructions as unavailable at launch, while supporting explicit emotion and pace tags. That distinction separates a request to describe an overall performance in ordinary language from a supported control placed inside the text. It also illustrates why the model announcement and the API’s current feature definitions need to be read together.

For deployment, the SDK can run in a server or desktop application, and browser applications obtain session credentials through their server. The phone connection uses SIP and is available to enterprise customers at launch. Saved agents have versioned configurations that can be reused across sessions. These details place the release closer to an application platform than a standalone transcription endpoint, with the surrounding software responsible for how each conversation connects to real systems.

Mistral is offering the new speech family through hosted services, with on-premises or private-cloud deployment discussed through its sales team. The TTS model card labels version 3.0 as a public preview. Some launch specifications differ between the announcement and model documentation, so the published material does not yet provide one consistent account of pricing and language coverage. Actual conversational performance will depend on the selected models, tools and integration, beyond the speech components alone.