Mistral’s Voxtral 3 joins listening, speaking and action

Three jobs behind one voice

The voice-agent documentation describes a chain: transcription turns speech into text and detects when the caller has finished; a general model reasons and calls tools; speech generation voices the answer. The SDK builds on LiveKit for audio transport. Developers still define what the agent knows and which actions it can take. Mistral’s original announcement presents the connected voice stack.

Access has boundaries. The Small Transcribe 3 model card labels the release a public preview. Browser and SDK routes are documented, while phone-line access through SIP is available to enterprise customers at launch through their account team. A demonstration of a booking is not evidence that a production service is ready.

Language coverage also belongs to the component. The current language guide lists 50 languages for file transcription, 45 languages and dialects for realtime transcription, and 35 languages with more than 50 dialects for speech generation. Those are different lists, not one promise that every capability works identically everywhere.

A good transcript still needs the right account number

Artificial Analysis’s non-streaming comparison gives Voxtral Small Transcribe 3 a 2.1% word error rate, between Microsoft’s MAI-Transcribe-2 at 2.0% and ElevenLabs’ Scribe v2 at 2.2%. The index combines roughly eight hours of agent-style speech, parliamentary speech and earnings calls. It measures transcription errors, not completed support requests; these small displayed gaps do not establish a decisive advantage for every language or recording. The evaluator’s launch report keeps transcription, streaming and speech-generation results separate.

BeSimple AI’s Voice Code Bench explains a different failure: an otherwise tidy transcript can corrupt the email address or reference number a system needs. Its 300 human-recorded English clips measure exact recovery of structured values as well as ordinary word errors. The published baseline table inspected here covers earlier Voxtral models, not a new independent Voxtral 3 result.

Mistral’s launch page separately presents a speech-generation evaluation using Voice Code Bench prompts, two voices per system and Gemini 3.8 Flash as judge. That vendor-presented test is not the same procedure as BeSimple’s human-recorded speech-to-text benchmark. The benchmark name alone does not make their scores interchangeable.

The bill and the pause have several parts

Component prices use different units. Mistral’s API price list charges transcription by audio minute and speech generation by character; the launch lists separate voice-agent rates. A low transcription rate is not the price of a complete conversation.

The latency guide makes a similarly useful distinction. A model’s processing time leaves out the network and application. For an agent, the relevant wait starts when the caller stops speaking and ends when the answer begins, including turn detection, reasoning, tools and speech. Our earlier reporting on support-call tests shows why a pleasant voice and a successful task need separate measures.

European processing needs the right route

Mistral offers European API hosting and a sales-led route to private deployment. Its regional-inference guide adds essential detail: the global endpoint makes no specific location commitment; dedicated regional inference carries a 10% surcharge, and control-plane data such as billing and account settings is not all covered by that geography.

The same guide currently excludes stateful Agents, Batch and Files features from regional endpoints. That is a reason to confirm the exact voice-agent configuration and contract, rather than assume the whole new stack inherits a regional guarantee. The launch links private deployment to its team; it does not supply downloadable Voxtral 3 weights with an open licence.

Voxtral 3 therefore offers a concrete new set of building blocks. The next useful evidence is a complete call: the correct details captured, the right tool action taken, interruptions handled and the actual wait and cost recorded in the intended deployment.