Voxtral 3
Speech models from Paris-based Mistral AI for transcription, speech generation and voice agents, used through the API.
What it does
Voxtral 3 is a family of Mistral AI models for spoken language. Voxtral Mini Transcribe 3 and Voxtral Small Transcribe 3 transcribe recorded files in more than 50 languages, with timestamps, custom vocabulary for specialist terms and – depending on the model – speaker labels. Voxtral Mini Transcribe Realtime 3 handles live audio in more than 35 languages and detects when a speaker has finished. Voxtral TTS 3 turns text into speech in more than 35 languages and 50 dialects, using preset voices or a recreated voice – from a short sample or professionally from around 30 minutes of audio – with control over emotion and pace. A Voice Agents SDK for Python and Node.js (plus a browser integration) combines listening, answering through a language model and speaking into one pipeline.
Who it suits
Development teams building call notes, subtitles or phone and voice assistants into their own applications. Without programming skills, you will hardly be able to use the models.
Data protection for businesses
Mistral AI is based in Paris. Using the api.eu.mistral.ai endpoint keeps requests in the EU; regional endpoints cost 10% extra. Account data, billing and usage analytics may still be processed outside the EU, and you should check beforehand whether all Voxtral 3 models are available there. Even on paid API access, inputs may be used for training unless you opt out of “Anonymous improvement data” under “Privacy” in the admin console. Stored voices are deleted after 30 days by default; the period can be set between 1 and 365 days. A data processing agreement is available on the legal pages. Zero data retention is available on request, only for pay-as-you-go and stateless calls; Mistral decides case by case. Deployments on your own infrastructure are available on request through sales.
Limits
The models are labelled public preview. Mistral does not mention a free allowance. Mistral does not state open weights for Voxtral 3. SIP telephony is available only to Enterprise customers at launch and, according to the documentation, only for inbound calls. Whether you may recreate someone’s voice is a legal question you need to settle yourself.