Voxtral 3 brings transcription, voice and real-time actions together

On 20 October 2026, Mistral introduced Voxtral 3, a family of voice models accompanied by an SDK for conversational agents. The proposal covers four functions that were often assembled separately: listening, interpreting, speaking and acting on external systems.

The range includes batch and real-time transcription models, multilingual speech synthesis and a layer that coordinates reasoning with tool calls. It can support phone service, internal assistants, accessibility, meetings and voice experiences at events.

The key difference is not just turning sound into text. An agent can detect when someone has finished speaking, check a booking system, confirm availability and answer aloud without leaving the conversation.

Voxtral Small Transcribe 3 supports more than 50 languages and adds diarisation to identify who said what. It also accepts keywords or prior context to improve recognition of names, brands, medical terms and technical vocabulary that general-purpose systems often get wrong.

The real-time version supports more than 35 languages and, according to Mistral, keeps latency below 500 milliseconds. Voxtral Mini TTS 3 generates speech in more than 35 languages and 50 dialects, with controls for emotion, pace and emphasis.

These figures come from the provider’s documentation and evaluations. A real deployment needs to be tested with the relevant audio, accents, noise, microphones and vocabulary; an average benchmark does not represent a conference room or a call with poor coverage.

Mistral offers access through an API hosted in Europe and the option to deploy components on an organisation’s own infrastructure. For teams handling sales conversations, internal meetings or customer recordings, where audio is processed and stored can matter as much as accuracy.

The published price for Voxtral Small Transcribe 3 is $0.004 per minute. The family includes a Mini variant aimed at lower cost and resource use, as well as real-time options. The model price is only one part of the bill: telephony, storage, review, observability and connected systems can cost more across the full service.

Local deployment does not automatically remove risk. Access controls, encryption, retention policies, activity logs and a clear separation between original audio, transcripts and derived data are still needed.

At an event, the same technology can create captions, identify speakers, produce summaries and answer questions about the agenda or venue. In business operations, it can log incidents by voice, check inventory, complete a form or assist a technician who needs to keep both hands free.

The SDK includes tool calls, but a spoken conversation should not turn every request into authorisation. Looking up information, preparing a proposal and taking an irreversible action are different levels. Payments, booking changes, access to sensitive data and external messages need limited permissions and explicit confirmation points.

It is also important to plan for cases where the system does not understand, the network degrades or a tool fails. A reliable voice experience includes repetition, clarification, transfer to a person and a visual alternative, not just a low-latency demo.

Expressive synthesis and voice adaptation open creative possibilities, but they also increase impersonation risk. A person’s voice should not be used without verifiable permission, and audiences need to know when they are speaking with an automated system or hearing generated content.

In settings with several people, recording and diarisation affect everyone in the room. Before capturing audio, organisations should define notices, a lawful basis, retention periods and channels for people to exercise their rights.

Voxtral 3 shortens the technical distance between an idea and a working voice agent. The remaining challenge is organisational: choose a useful process, test it with real audio and set precise limits on what the system can do after listening.