Mistral AI Releases Voxtral 3: Speech-to-Text, TTS and a Voice Agent SDK

Mistral released Voxtral 3, a family of three transcription models, a text-to-speech model, and a voice agent SDK built on LiveKit, all in public preview on its API. The pitch is one place to listen, reason, call tools and speak, with each turn completing in under a second and the agent stopping when the caller interrupts.

The pricing is aggressive: batch transcription runs $0.0035 to $0.004 per minute, speech costs $13 per million characters, and the full voice agent is $0.04 per minute with Mistral Small 4. On Artificial Analysis, Voxtral Small Transcribe 3 posts 2.1% word error rate, third on that board. Batch transcription covers 50 languages, including the 24 official EU languages.

Voice is where a lot of practical AI will actually meet people: support lines, scheduling, accessibility, transcription for meetings and medical notes. A credible, low-cost stack from a European lab, with EU endpoints and zero-retention options listed on OpenRouter, gives buyers a real alternative to the US majors. We think that competition is healthy, and the per-minute prices make voice agents affordable for small businesses, not just call-center giants.

The numbers deserve a careful read. Most accuracy benchmarks here are self-reported by Mistral. The independent Artificial Analysis ranking is good but not first, and speed is a clear gap: Small's speed factor trails Microsoft's MAI-Transcribe-2 by a wide margin. The realtime model, the one a live voice agent depends on, scores 7.4% word error at its fastest 240 ms setting, against 4.2% for Gemini 3.5 Realtime. On a phone line, that difference shows up as misheard names and wrong account numbers.

We also note what did not ship: open weights. Mistral built part of its reputation on releasing models people could run themselves. This launch is API-first, with self-hosting through sales.

The harder issue is voice cloning. Pro Voice adapts a speaker's style from 30 minutes of audio, and the speech model offers explicit emotion and emphasis controls. Those are useful for brands and accessibility, and they are also exactly what makes phone scams and impersonation cheaper. As voice agents get faster and more natural, the people on the other end deserve to know when they are talking to software.

I'm watching realtime accuracy improvements, any open-weight release, and what consent and disclosure controls Mistral puts around voice adaptation.

Speech has become a crowded race, with Microsoft, Google, ElevenLabs and OpenAI all shipping transcription and speech models. Mistral is competing on price, diarization and multilingual European coverage.