Mistral launches Voxtral 3 for voice agents with European hosting

Mistral AI has launched Voxtral 3, a new family of speech recognition and speech synthesis models, together with a toolkit for building voice agents. The agents can take phone calls, work out what the caller wants, look things up in internal systems and reply in a natural voice. The models run in Mistral's API with European endpoints or on the customer's own infrastructure. For Norwegian organisations with contact centres or switchboards, that means a European alternative to US speech services, but also new questions about voice cloning and consent.

Mistral published the announcement on its own blog on 20 October 2026 and posted it on X the same day at 18:00 Norwegian time. The company says Voxtral 3 is available from today.

Four models and one toolkit

Voxtral 3 consists of two kinds of models. Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 turn speech into text. Mistral states support for more than 50 languages, including all 24 official EU languages. Norwegian is on the list of languages the company tested. The models tell different speakers in the same recording apart and can be given a list of technical terms or names to recognise them more reliably. The Mini variant has four billion parameters.

Voxtral Mini Transcribe Realtime 3 is built for real time. Mistral says it transcribes with under half a second of delay and detects when a person has finished speaking. That is meant to let the agent answer without unnecessary waiting and stop talking when the caller interrupts.

Voxtral Speech 3, also referred to as Voxtral Mini TTS 3, turns text into speech. Mistral states support for more than 35 languages and more than 50 dialects. The model adapts emotion, pace and emphasis to the conversation, and developers can control this line by line with their own tags. Preset voices are available in 18 languages, including Norwegian.

The new voice agent toolkit is built on the LiveKit framework. It connects speech recognition, a language model that reasons and calls tools, and speech synthesis in one stream. Mistral says each turn in the conversation takes under one second. The agent can place and answer calls over SIP, run in browsers and apps over WebRTC and connect to existing telephony and contact-centre software.

Prices and the company's own measurements

Mistral has published prices. Voxtral Mini Transcribe 3 costs 0.0035 dollars per minute in batch, Voxtral Small Transcribe 3 costs 0.004 dollars per minute, and the realtime model costs 0.008 dollars per minute. Speech synthesis costs 13 dollars per million characters. No price is given for running the stack on your own infrastructure; customers have to contact Mistral.

The company writes that Voxtral Small Transcribe 3 beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription. Mistral also says its speech synthesis costs less to run than comparable models such as OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Both claims rest on measurements Mistral itself selected and published, and several use a language model as the judge. They have not been independently verified.

Voice cloning from 30 minutes of audio

The most sensitive feature is called Pro Voice. With up to 30 minutes of recorded audio, Voxtral can reproduce a specific person's voice, expression and speaking style across all supported languages. Mistral aims the feature at professional users, for example organisations that want their own brand voice.

The announcement does not say how Mistral ensures that the person behind the voice has consented, or which terms apply to the use of a cloned voice. Nor does it say whether the models will be released with open weights.

Data in Europe or on your own servers

Mistral says Voxtral 3 is offered through European, US and global endpoints in its API. Organisations can also run the whole stack on their own infrastructure or in a private cloud. According to Mistral, customers can then decide where audio is processed, how data is handled and whether conversations are retained. How long audio is kept by default in Mistral's own API is not stated in the announcement.

What Norwegian leaders should note

• Contact centres and switchboards considering voice agents get a European vendor with published prices and an option to self-host. That can reduce dependence on US cloud services, but Norwegian quality should be tested on your own calls and dialects before signing.
• Phone calls often contain personal data. Settle the data processing agreement, the choice of endpoint and the retention period before sending audio, since the announcement does not state default retention.
• Voice cloning needs its own rules. Decide who can consent to an employee's voice being used, and how customers are told they are talking to a machine.
• Per-minute prices make it possible to cost a pilot. Remember that the language model doing the reasoning comes on top.

Sources and media

• Mistral AI: Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages, blog post dated 20.10.2026
• Mistral AI on X: Introducing Voxtral 3, post published 20.10.2026 at 18:00 Norwegian time
• Sources were checked on 20.10.2026.
• Illustration: AI-generated with Grok Bot / hogby.ai