TipMistral Voxtral 3: European speech AI that understands the Nordic languages, and an hour-long meeting costs about 21 cents to transcribe
France’s Mistral AI has released Voxtral 3, a new family of AI models for speech. They transcribe audio in more than 50 languages, including Swedish, Danish, Finnish and Norwegian, read text aloud with emotion and can power voice agents that answer the phone. Everything can run in the EU or on your own servers. Here is how your business can use it, and what to watch out for.
On 20 October 2026 Mistral AI released Voxtral 3. It is four models, all in Public Preview, meaning a preview version that may change: three for turning speech into text and one for turning text into speech. At the same time Mistral released a tool for building voice agents, that is AI that can hold a conversation, for example on the phone. Voxtral 3 is used through Mistral’s API, so it is primarily for businesses and developers.
Transcription covers more than 50 languages, including the EU’s 24 official languages. Swedish, Danish, Finnish, Norwegian and Icelandic are included in Mistral’s language test. The model can tell who is saying what, and you can give it a list of names and technical terms so they are spelled correctly. Mistral calls this context biasing. For live transcription in real time, Mistral promises responses in under half a second in more than 35 languages.
Prices are low. Voxtral Mini Transcribe 3 costs 0.0035 dollars per minute, so an hour-long meeting comes to about 0.21 dollars. The larger Voxtral Small Transcribe 3 costs 0.0040 dollars per minute and the real-time version 0.008 dollars per minute. The speech synthesis model Voxtral Mini TTS 3 costs 13 dollars per million characters, and a voice agent costs 0.04 dollars per minute with the Mistral Small 4 model and 0.08 dollars with a larger model.
Speech synthesis covers more than 35 languages and 50 dialects and adapts emotion, pace and emphasis to the context. Ready-made voices are available in 18 languages. Of the Nordic languages only Norwegian is among them, and Mistral lists no ready-made Swedish voices. The Pro Voice feature can imitate a person’s voice from 30 minutes of audio.
Why does it matter in the Nordics? Mistral is a European company, and you can choose to run Voxtral through Mistral’s API in the EU or on your own servers. Then you decide where the audio is processed and whether conversations are stored. For health clinics, law firms and others who handle sensitive conversations, that can matter more than getting the very best result.
How to get started
- Start with transcription. It is the cheapest and easiest to test: record an internal meeting and compare with the tool you use today.
- Make a glossary of names, places and technical terms and send it along. Our prompt Make a glossary for better transcription helps you.
- Always tell participants that the meeting is being recorded and what the recording will be used for, before you press record.
- Only imitate voices with the person’s explicit permission, and never to deceive anyone.
- It is a preview version: check prices and terms in Mistral’s documentation before you build something customers will use.
More on TipAI
Sources
- Mistral AI: Voxtral 3 – Voice Agent SDK & Speech Models in 50+ Languages (priser och språk) (retrieved 2026-10-25)
- Mistral Docs: Changelog, 20 oktober 2026 (Voxtral 3, fyra modeller i Public Preview) (retrieved 2026-10-25)
- The Decoder: Mistral AI launches Voxtral 3, a full voice agent stack (20 okt 2026) (retrieved 2026-10-25)
Free cheat sheet: Get started with AI
A printable page with the basic formula, five things to try and how to use AI safely. Enter your email and we will send it straight away – free, unsubscribe any time.
Get the cheat sheet free →TipAI explains AI in plain language for beginners, families and small businesses. Every article has sources and a "How to get started" section.