Mistral's Voxtral 3 is a full voice-agent stack: it listens, thinks, calls tools and talks back

Mistral released Voxtral 3 on October 20. It is three things at once: new speech-to-text models, a new text-to-speech model, and a voice agent SDK that wires them to a language model so an agent can hold a phone or web conversation and do things during it. It is available today through the Mistral API, on European, US and global endpoints, and in Mistral's Studio playgrounds. Mistral says it can also be self-deployed on your own infrastructure.

What it hears. Voxtral Small Transcribe 3 is the top transcription model. Mistral says it handles 50 languages, including the 24 official EU languages, labels who said what (diarization), takes a list of key terms or free text to get names and jargon right (context biasing), and detects when a speaker has finished so an agent can answer without a long pause. Mistral says it beats Whisper, ElevenLabs Scribe v2 and Gemini 3.5 Transcribe on its charts. Voxtral Mini Transcribe 3 is a 4-billion-parameter version, and Voxtral Mini Transcribe Realtime 3 is the streaming one, which Mistral says transcribes at under 500 milliseconds of latency.

What it says. Voxtral Speech 3 (called Voxtral Mini TTS 3 in places on the same page) reads emotion, pace and emphasis from the conversation on its own, and you can also set them line by line with tags like [emotion=sad][pace=slow]. It covers 35+ languages and 50+ dialects, with preset voices in 18 languages, and it takes phoneme hints for words it would otherwise mispronounce, such as product names. A new Pro Voice feature builds a custom voice from about 30 minutes of recorded audio.

The agent part. The SDK, built on the open-source LiveKit framework, runs the realtime transcriber, a general model such as Mistral Small for reasoning and tool calls, and Voxtral Speech 3 as one streaming pipeline. Mistral says no stage waits for the one before it, each turn finishes in under a second, and the agent stops talking when the caller cuts in. It can use function calling, web search and retrieval over your own systems, take and place calls over SIP, and run in browsers and apps over WebRTC. Because each stage is a separate model, you can swap the language model, the voice or the language without rewriting the agent.

Prices, from Mistral's table.

• Voxtral Mini Transcribe 3 (batch): $0.0035 per minute
• Voxtral Mini Transcribe Realtime 3: $0.008 per minute
• Voxtral Small Transcribe 3: $0.0040 per minute
• Voxtral Speech 3 (batch and realtime): $13 per million characters

Who it's for: teams building phone support lines, booking and service-desk agents, or hands-free assistants who want one vendor for listening, speaking and tool use, especially in Europe, where keeping audio in the EU matters.

The caveats. The text-to-speech charts are Mistral's own evaluations, scored by another AI model (Mistral lists "gemini-3.8-flash" as the judge) on prompt sets "selected for release," and the sub-one-second turn is Mistral's claim. Run your own calls, with your own accents and product names, before you trust the curves. On self-deployment: when we checked Mistral's Hugging Face page on the morning of October 20, it listed earlier Voxtral models but no Voxtral 3 weights, so ask Mistral how self-hosting is licensed before you plan on it. And if you use Pro Voice, clone only voices you have permission to use.

We haven't tested Voxtral 3 ourselves; this is based on Mistral's announcement page and thread, read on October 20. Not sponsored, no affiliate links.