Introduction

Mistral just gave its voice lineup a full refresh. Voxtral 3, announced on October 20 in Mistral's launch post, is a new family of speech-to-text and text-to-speech models plus a voice agent SDK that strings listening, thinking, tool calls and speaking into one real-time pipeline. It's live today through the Mistral API and the Studio playgrounds, with transcription starting at $0.0035 a minute. This is a closed, hosted release, with no new open weights on Hugging Face, but if you build talking characters, narrated content or voice assistants, it's a big new option.

What Shipped

Four new models

All four are listed as Public Preview in Mistral's docs, dated October 20:

  • Voxtral Small Transcribe 3 (voxtral-small-transcribe-3-0): the flagship transcription model, with 50+ languages, speaker diarization (who said what), context biasing for names and jargon, and end-of-turn detection. $0.004 per minute.
  • Voxtral Mini Transcribe 3: a 4B-parameter batch transcription model. $0.0035 per minute.
  • Voxtral Mini Transcribe Realtime 3 (voxtral-mini-transcribe-realtime-3-0): streaming transcription over WebSocket with latency you can set from 80 ms to 2,400 ms, plus diarization and end-of-turn detection. $0.008 per minute.
  • Voxtral TTS 3 (voxtral-mini-tts-3-0, also called Voxtral Speech 3): expressive text-to-speech in batch and realtime from the same API.

The voice side

This is the part I care about most. Voxtral TTS 3 reads the meaning of your text to pick its emotion and pacing, and it also takes explicit instructions for emotion, pacing and emphasis. The model page lists 34 emotion tags and 2 pace tags, streaming with about 90 ms to first audio, and five output formats. You can give it a phoneme hint for a made-up or rare word, so a character named "Xyra" or "Quivello" comes out the way you meant it.

Voice cloning comes in two levels. There's zero-shot cloning, and there's Pro Voice, which Mistral says captures a speaker's character, emotional range and speaking style across languages from about 30 minutes of audio.

Two numbers didn't match when I checked. The launch post says Voxtral Speech 3 costs $13 per million characters and covers 35+ languages and 50+ dialects, while the model page showed $16 per million characters and 9 languages. Check the model page before you budget a big narration job.

The voice agent SDK

The new SDK is built on LiveKit. Realtime transcription listens and spots when the caller is done talking, a Mistral language model reasons and calls your tools, and Voxtral speech answers in a tone that fits the conversation. Mistral says each stage streams into the next, so a turn finishes in under a second, and you can swap the LLM, the voice or the language without rewriting the agent. Organizations that need it can talk to Mistral about running the whole stack on their own servers.

Early arena results

placed Voxtral Small Transcribe 3 at #3 on its Speech to Text leaderboard, and Voxtral 3 TTS entered the top 12 in both of its Text to Speech arenas.
has added Voxtral TTS 3 as well. Mistral's own charts claim Small Transcribe 3 beats Whisper large-v3, Scribe v2 and Gemini 3.5 Transcribe on multilingual word error rate, but those are vendor numbers.

Try It Yourself

The quickest way to hear it is the speech playground in Mistral Studio, which also covers transcription, voice cloning and voice agents. For your own code, text-to-speech and cloning use the /v1/audio/speech endpoint and transcription uses /v1/audio/transcriptions.

Lines that test the expressive side and the pronunciation hints:

Wait. This painting came from Reims? It's the one my grandmother searched for fifty years!
Welcome aboard. I'm Xyra, your personal assistant, and the stars are right on schedule.
Let's end the debate here: it's GIF.
Your delivery window is 9:30 to 11:30. Shall I text you the details?

My tip: write the emotion into the line itself first, since the model reads context, and add explicit emotion or pace tags only when it guesses wrong.

Why It Matters for Makers

Voxtral 3 puts expressive TTS, cheap multilingual transcription and a ready-made voice agent pipeline under one API key. Creators can use it for audiobook-style narration with directed emotion, multilingual dubbing drafts, interactive characters that talk back, and podcast or video transcripts with speaker labels. Transcription at well under a cent a minute makes subtitling a back catalog cheap.

The trade-offs: these are API models, with self-hosting only through a conversation with Mistral's team. Everything is labeled Public Preview, the TTS price and language count differ between the launch post and the docs, and the benchmark charts come from Mistral. Pro Voice needs about half an hour of clean audio. As always, only clone voices you have permission to use.

Conclusion

Voxtral 3 brings Mistral four new speech models, expressive TTS with emotion and phoneme control, transcription in 50+ languages from $0.0035 a minute, and a LiveKit-based SDK for voice agents, all live today on the Mistral API. Open the Studio playground, hand it a dramatic line, and see how it handles the feeling. Happy voice-making!

Original source:

https://mistral.ai/news/voxtral-3/

Model pages: Voxtral TTS 3 · Voxtral Small Transcribe 3 · Voxtral Mini Transcribe Realtime 3 · Announcement:

—Anabel ♡