Voxtral 3: Mistral builds a full tech stack for voice agents
Mistral's Voxtral 3 brings speech-to-text and text-to-speech models plus a voice agent SDK, from $0.0035 per minute. More on what it can do and what it costs.
This article in other languages:العربيةDeutschEspañolFrançaisItaliano日本語Polski简体中文
In this article
On October 20, Mistral released Voxtral 3, a new family of speech-to-text and text-to-speech models, together with an SDK for building voice agents. Everything is available through the Mistral API, and the announcement collected almost 500 points in its Hacker News thread.
What’s new in Voxtral 3?
According to the announcement, Voxtral 3 is four models and one SDK:
- Voxtral Small Transcribe 3 – the most accurate transcription model, with speaker recognition (so-called “diarization”), domain-term hints and end-of-turn detection.
- Voxtral Mini Transcribe 3 – a smaller model (4B parameters) for transcription.
- Voxtral Mini Transcribe Realtime 3 – live transcription with latency under 500 ms, designed for agents.
- Voxtral Speech 3 – speech synthesis with control over emotion, pace and emphasis. According to the model documentation, the first audio arrives after about 90 ms.
On top of that there’s a voice agent SDK built on LiveKit. Listening, reasoning, tool calling and speaking run in a single stream, and according to Mistral each conversation turn is processed in under a second. The agent can use web search and data from our own systems, and you can connect to it over SIP (that is, an ordinary phone) or WebRTC.
Voxtral Speech 3 can also clone voices. Basic cloning needs more than 30 minutes of recordings, and for professional use Mistral offers so-called “Pro Voice”, which according to the announcement also captures a speaker’s distinctive style.
How much does Voxtral 3 cost?
Prices from the announcement:
| Model | Price |
|---|---|
| Voxtral Small Transcribe 3 | $0.004 per minute |
| Voxtral Mini Transcribe 3 | $0.0035 per minute |
| Voxtral Mini Transcribe Realtime 3 | $0.008 per minute |
| Voxtral Speech 3 | $13 per 1M characters |
| Voice agent API | $0.04 per minute (with Mistral Small 4) or $0.08 per minute (with another model) |
So an hour of transcription with the smallest model comes to about 21 cents. Mistral compares itself here with Whisper, ElevenLabs, Deepgram and Gemini, but these are only the company’s own measurements, so I’ll hold off on judging until independent tests arrive.
Which languages does Voxtral 3 support?
Plenty, at least for transcription. According to the announcement, the speech-to-text models support more than 50 languages, including all 24 official EU languages, as measured on the FLEURS benchmark.
Speech synthesis is less clear. The announcement mentions more than 35 languages and 50 dialects in the voice library, but the model card in the documentation lists 9 languages (check the voice for your language in the playground before you build anything on it).
Is anything wrong?
Yes, and that’s exactly what dominates the Hacker News discussion. Earlier Voxtral models had open weights (e.g. Voxtral Mini 3B on Hugging Face), but the Voxtral 3 announcement mentions neither weights nor a license. Mistral writes about “self-deployment on your own infrastructure”, but an on-premise deployment requires contacting sales. One commenter summed it up briefly: “Closed model unfortunately”.
Quality reviews are mixed too. Some people praise the transcription and speaker recognition in previous versions, others write that the new voices in the demo sound more artificial, and for local transcription they recommend Parakeet. So let’s take the announcement with a grain of salt until we see broader tests.
How to try Voxtral 3?
We have two options:
- Playgrounds in Mistral Studio – you can test transcription, speech synthesis, voice cloning and voice agents there after logging in to the Mistral console.
- API – we generate a key in the console, and the details are in the audio documentation. Endpoints are available in the EU, the US and globally.
If we’re looking for something to run on our own machine, Voxtral 3 won’t do (at least for now). In that case it’s better to reach for the older, open versions and run AI locally. And anyone building a fast voice agent can compare Voxtral 3 with Mercury Voice, which Inception showed together with Mercury 3.
FAQ
Is Voxtral 3 open source?
Nothing suggests so for now. The announcement gives no license or links to weights, and deploying on your own infrastructure requires contacting Mistral. The older Voxtral models have open weights.
How much does transcription cost in Voxtral 3?
From $0.0035 per minute (Voxtral Mini Transcribe 3) to $0.008 per minute in live mode. Speech synthesis costs $13 per million characters.
How is Voxtral 3 different from the previous Voxtral?
Previous versions were separate models for transcription or speech synthesis, largely with open weights. Voxtral 3 is a whole family of models plus a voice agent SDK, available through the API.