Mistral releases Voxtral 3 to connect live voice conversations with business tools

Mistral AI released Voxtral 3 on October 20, bringing speech recognition, reasoning, connected software tools and spoken replies into one live conversation pipeline. The speech-model family and developer toolkit are available through Mistral’s EU-hosted API or on customers’ own infrastructure, targeting voice agents that can complete tasks rather than only answer questions.

From a caller’s correction to a booking

The voice-agent SDK—a software development kit—connects listening, reasoning, tool calling and speaking in a streaming process. Tool calls let an agent retrieve information or take action through connected business software during a phone or web conversation. Mistral positions the system for contact centers, service desks, virtual assistants and hands-free copilots.

Mistral illustrates that workflow with a medical-booking example. A caller corrects a reference number, and the agent uses the corrected number to retrieve a French referral. It explains the referral in English, offers an appointment and books it after the caller agrees. This is the company’s demonstration of a multi-step exchange, not evidence of performance across real medical calls.

The pipeline is modular rather than a single model doing every job. Mistral says developers can change the language model, voice or language without rewriting the agent. That separates the choice of how an agent reasons from how it listens and speaks.

Listening includes knowing when to wait

Voxtral Small Transcribe 3 covers more than 50 languages. Its features include diarization, which identifies who spoke, and context biasing, which uses supplied terms to improve recognition of names or specialist vocabulary. Mistral says developers can supply key terms or free-form text without fine-tuning the models. The Mini transcription variant has 4 billion parameters and costs $0.0035 per audio minute.

For live applications, Mistral says Voxtral Realtime Transcribe 3 delivers transcription with less than 500 milliseconds of delay in more than 35 languages. That figure describes transcription, not completion of an entire spoken request. The model also uses sound and meaning to predict when a person has finished speaking, aiming to respond quickly without cutting the caller off.

Mistral’s published tests also show a speed-versus-accuracy tradeoff. On its 35-language FLEURS comparison, the realtime Mini model’s reported error rate falls as the allowed delay increases. The measure uses word errors for most languages and character errors for Japanese, Korean and Chinese. These are company-published benchmark results, not a measured success rate for finished voice agents.

7.4%At 240 milliseconds

Mistral reports a 7.4% error rate for Voxtral Mini Realtime 3 at the 240-millisecond setting on its FLEURS 35 comparison.

5.7%At 2,400 milliseconds

At the 2,400-millisecond setting, Mistral reports a 5.7% error rate on the same comparison.

A directed voice, with control over the audio

On the speaking side, Voxtral Mini TTS 3 turns text into audio in batch or realtime, supporting more than 35 languages and 50 dialects. Mistral prices it at $13 per million characters. The company says the model can infer emotional delivery from surrounding text, while explicit instructions let developers adjust emotion, pacing and emphasis.

Its Pro Voice adaptation targets a more specific identity than a preset voice. Mistral says 30 minutes of audio can reproduce a speaker’s distinctive vocal character, emotional range and speaking style across languages. The feature is aimed at professional uses where basic voice cloning is insufficient.

Deployment is also part of the release’s pitch. Mistral says organizations can control where audio is processed, how it is handled and whether conversations are retained. Self-deployment lets them keep sensitive conversations within their own environment; the hosted alternative runs through Mistral’s API in Europe.