Mistral Voxtral 3: speech-to-text in 50+ languages and text-to-speech in 35+

What's new

Mistral AI launched Voxtral 3 on October 20, 2026, according to The Decoder. It is a full family of speech models plus a so-called Voice Agent SDK, a toolkit for developers who want to build voice assistants.

The family has four parts: Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 convert speech to text in 50+ languages (Mini is for batch, meaning finished recordings). Voxtral Mini Transcribe Realtime 3 does it live in 40+ languages with latency under 500 milliseconds. Voxtral Speech 3 / Mini TTS 3 turns text into speech in 35+ languages and 50+ dialects, and 18 of the languages have ready-made voices to choose from. According to Mistral, the transcription covers all 24 official EU languages.

What's clever

The SDK combines everything in one stream: listening, thinking, using tools and responding aloud. According to The Decoder, it is built on LiveKit. That means an agent can talk to external systems in the middle of a conversation, for example to book an appointment or check an invoice.

According to Mistral, the system can sense when a person has finished speaking (both from the audio and from the meaning), tell speakers apart (diarization) and remember specialized terms via context biasing, so industry-specific words are spelled correctly. It also supports function calling, web search and phone and web calls via SIP and WebRTC. Conversations can be wrapped up in under one second.

Cheaper or better?

Prices are listed in dollars: Mini Transcribe 3 costs $0.0035 per minute, Small Transcribe 3 costs $0.004 and the realtime model costs $0.008 per minute. Text-to-speech costs $13 per million characters. The voice agent API costs $0.04 per minute with Mistral Small 4 and $0.08 with other models. No monthly price has been disclosed for ordinary users.

On quality, Mistral says Voxtral Small has a 4.1% word error rate on the FLEURS 45 test. By comparison, Whisper large-v3 has 16.0%, Scribe v2 from ElevenLabs has 6.3% and Gemini 3.5 has 6.1%. The realtime model (480 ms) reaches 5.8% on the Artificial Analysis test. Mistral also claims that its text-to-speech offers a better balance of price and quality than ElevenLabs, OpenAI and Deepgram. All figures are Mistral's own and have not been independently verified.

What it's good for

The models suit companies that want phone-based customer service, meeting notes with multiple speakers, or automatic booking and invoice checks by voice. Because several EU languages are included, it may be relevant for European companies.

For those who care about data control, there is an EU-hosted API and a US endpoint. The models can also run on the customer's own infrastructure, on-premises or in a private cloud, by contacting Mistral's sales team. The sources do not say whether the models are open source or whether everything is available in Denmark.