Mistral Voxtral 3: speech models and a voice agent SDK for 50+ languages

What this is about

On October 20, 2026, Mistral AI introduced Voxtral 3: a family of speech-to-text and text-to-speech models, plus an SDK (developer toolkit) for voice agents. It targets call centers, service desks, virtual assistants and hands-free copilots. Voxtral 3 is available through Mistral's API with European, US and global endpoints, in the Studio playgrounds for testing, and for self-deployment. The source doesn't say whether the weights can be freely downloaded. For on-premises or private cloud deployment, you're told to contact the Mistral team.

Mistral argues that a reliable voice agent has to do four things at once: listen accurately (languages, accents, speakers, interruptions), understand context, speak naturally and with emotion, and take multi-step actions through connected systems.

Speech recognition: 50+ languages

Transcription now covers 50 languages, including the 24 official EU languages and many Asian and Indian languages with little training data. Key features:

• Diarization: Identifying who spoke when.
• Context biasing: You pass in key terms or free text to improve recognition of technical vocabulary, with no fine-tuning.
• End-of-turn detection: Acoustic and content signals show when the speaker has finished, so the agent answers without delay and, ideally, doesn't interrupt.

According to Mistral, Voxtral Small Transcribe 3 achieves top results in diarization and multilingual transcription, including in English, French and Spanish, and beats Whisper, Scribe v2 and Gemini 3.5 Transcribe. These are vendor claims, and the text gives no figures. The Mini variant has 4B parameters. The realtime model delivers under 500 ms latency in more than 35 languages (elsewhere on the page: 40+), also with diarization and context biasing.

Speech output with emotion

Voxtral Mini TTS 3 comes in a batch and a realtime variant. It adapts emotion, pace and emphasis to the context of the text, with no extra instructions. It supports 35+ languages and 50+ dialects, and it accounts for differences in pronunciation, formality and spelling, as well as conventions for time, date, temperature, addresses and currencies. Voice presets are available in 18 languages and variants, including Arabic, Brazilian Portuguese, Catalan, Mandarin, Dutch, German, Greek, Hindi, Korean and Vietnamese. Other features include voice cloning from 30+ minutes of audio, instructions for emotion, pace and emphasis, contextual TTS and a vocabulary library. Pro Voice is meant to reproduce a speaker's character, emotional range and speaking style across languages, with just 30 minutes of audio, according to Mistral. Mistral says its price is below comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.

Voice agent SDK

The SDK is built on LiveKit and links listening, thinking and speaking into one streaming pipeline. The realtime transcription model detects the end of a sentence, a general-purpose model such as Mistral Small thinks and calls tools, and Voxtral TTS replies in the right tone. No stage waits for the one before it. Each conversational turn is meant to finish in under a second, and the agent stops talking when the caller cuts in. During a call, it can verify identity, look up invoices or book appointments, using function calling, web search and retrieval over your own systems. It places and receives calls over SIP, runs in browsers and apps via WebRTC, and can be connected to telephony, IVR and contact-center software. Because each stage is its own model, you can swap out the LLM, the voice or the language.

Pricing

Our take

Voxtral 3 is interesting for companies building voice agents that want data sovereignty: EU hosting or their own infrastructure, plus control over whether audio is stored. The fact that the SDK builds on LiveKit and the stages are swappable reduces vendor lock-in. The comparisons with Whisper, Scribe v2, Gemini and the TTS competitors are vendor claims, with no figures published in the text. You should check the docs on the inconsistencies in the source (realtime in 35+ vs. 40+ languages, voice cloning from 30 minutes). It's also unclear which parts are available as open weights.