Mistral launches Voxtral 3, a voice AI system for 50+ languages

French AI company Mistral has released Voxtral 3, a new family of speech models and a voice agent software development kit (SDK) designed to build real-time conversational AI systems that can listen, reason, speak, and take action. The transcription component covers 50+ languages, while the text-to-speech component supports 35+ languages and 50+ dialects.

The release includes two core speech models alongside a toolkit for developers. The transcription model, called Voxtral Small Transcribe 3, handles speech-to-text duties. The text-to-speech model, Voxtral Mini TTS 3, generates natural-sounding voice output that can shift its emotion, pace, and emphasis depending on the conversation.

Mistral says the Small transcription model reaches state-of-the-art results on multilingual speech recognition and diarization — the task of identifying who spoke and when. It also outperforms earlier models from OpenAI (Whisper), Google (Gemini 3.5 Transcribe) and Scribe (Scribe v2) on benchmark tests, according to the company ¹. The Mini variant, which contains 4 billion parameters, is priced at $0.0035 per minute of audio .

A key feature is context biasing, which lets developers feed the model key terms or technical vocabulary before transcription begins. This helps the system recognize domain-specific words without needing to be retrained, Mistral says ¹.

The voice agent SDK is what ties the whole system together. It combines listening, reasoning through a connected language model, calling tools such as calendars or databases, and speaking back to the user — all in a single streaming pipeline that runs in real time. Mistral positions the system for use in contact centers, virtual assistants, and hands-free copilots ¹.

For professional voice cloning, Mistral introduced Pro Voice, which can reproduce a speaker's distinctive tone and emotional range across different languages using just 30 minutes of recorded audio ¹.

As with other Mistral products, Voxtral 3 is offered through the company's European Union-hosted API or can be self-deployed on a company's own servers. This gives organizations control over where audio data is processed and whether conversations are stored ¹.

Mistral, founded in 2023 by former Meta researchers, has grown into one of Europe's leading AI companies. It counts ASML, the Dutch semiconductor equipment maker, among its investors ². The company's previous product lineup included large language models such as Mistral Large and the Mistral OCR system, but Voxtral 3 marks its first major push into dedicated speech and voice agent technology .

Corrections and updates

• Corrected on 12 November 2026. An earlier version said Voxtral 3 could act across more than 50 languages, but the source clarified that only the transcription component covers 50+ languages while TTS supports 35+ languages and 50+ dialects. An earlier version also cited a $24.7 billion valuation for Mistral, which no source supported; the revised version omitted that unverified claim. An earlier version claimed both transcription and speech synthesis supported 50+ languages, but the source states transcription supports 50+ languages while TTS specifically supports 35+ languages and 50+ dialects.

Key facts

• - Voxtral 3 supports transcription in 50+ languages; speech synthesis covers 35+ languages and 50+ dialects
• The Small Transcribe model beats earlier systems from OpenAI, Google and Scribe on benchmarks
• Context biasing lets the model handle technical vocabulary without retraining
• Pro Voice cloning works with just 30 minutes of audio
• Available via Mistral's EU-hosted API or self-deployment

Sources

1. Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages
2. Mistral AI — en.wikipedia.org

Watch