French AI company Mistral has released Voxtral 3, a family of four speech models that can transcribe audio into text, do it live as people talk, and turn written text back into speech. The launch, announced around October 20, also includes a kit for building voice assistants. Mistral says the models are available now through its developer platform.
The four models
- Voxtral Small Transcribe 3 is Mistral’s most accurate speech-to-text model. It supports many languages and “diarization”, which means labelling who said what in a recording.
- Voxtral Mini Transcribe 3 is a smaller, quicker version for processing recordings in bulk.
- Voxtral Mini Transcribe Realtime 3 transcribes live speech with under half a second of delay, according to Mistral, and can tell when a speaker has finished their turn. That is key for voice assistants that need to know when to reply.
- Voxtral Mini TTS 3 is a text-to-speech model that Mistral says adjusts emotion, pacing and emphasis to fit what it is saying. It also supports voice cloning.
Mistral says transcription works in more than 50 languages, including all 24 official EU languages and several less widely supported Asian and Indian languages. Text-to-speech covers more than 35 languages.
Building voice assistants
Alongside the models, Mistral released a voice agent toolkit built on LiveKit, an open-source platform for real-time audio. It links listening, thinking, using tools and speaking into one continuous flow. Because each stage is a separate model, a business can swap out the voice, the language or the AI “brain” without rebuilding everything.
What it costs
Voxtral 3 is aimed at developers and businesses and is billed by use. Mistral’s listed prices include:
- $0.0035 per minute for Mini batch transcription
- $0.004 per minute for Small transcription
- $0.008 per minute for real-time transcription
- $13 per million characters for text-to-speech
- $0.04 to $0.08 per minute for the full voice agent service, depending on which AI model powers it
At those rates, transcribing an hour of audio with the cheapest model costs around 21 cents. Companies that want to run the models on their own servers are asked to contact Mistral.
Why it matters
Speech is becoming a key battleground. Customer service lines, meeting note-takers and voice assistants all depend on accurate transcription and natural-sounding voices. Mistral, Europe’s best-known AI lab, is pitching itself as an alternative to US providers, offering EU-based servers alongside US and global options.
Voice cloning features also raise familiar concerns about impersonation, and Mistral’s announcement focuses on business uses rather than consumer apps.
What it means for you
There is no new consumer app here, and Mistral’s announcement does not say Voxtral 3 is coming to its Le Chat chatbot. You are more likely to meet it indirectly: in a transcription app, a company’s phone line or a meeting tool built on Mistral’s models. If you are a developer or run a small business, you can test the models in Mistral’s online playground and pay only for what you use.
Sources
- Voxtral 3 — Mistral AI