Mistral's Voxtral 3 bundles speech models with a sub-second voice agent SDK

Mistral's Voxtral 3 is a set of speech models paired with a voice agent SDK that stitches listening, reasoning, tool calling, and speaking into one real-time pipeline. On the input side, the transcription models cover more than 50 languages with speaker diarization and context biasing for domain vocabulary, and a realtime variant targets sub-500ms latency across 35 or more languages. On the output side, the text-to-speech model handles 35+ languages and 50+ dialects with controls for emotion, pacing, and emphasis, plus voice cloning from about 30 minutes of audio. The SDK is built on LiveKit, runs over SIP and WebRTC, and supports function calling and retrieval over your own data, with Mistral claiming each turn finishes in under a second.

The pricing is the clearest part of the pitch. Batch transcription starts at $0.0035 per minute, the realtime model is $0.008, text-to-speech is $13 per million characters, and the voice agent API is $0.04 per minute when paired with Mistral Small 4 for reasoning, or $0.08 with other models. The benchmark claims are vendor-reported, so read them with care: Mistral says its Small Transcribe 3 leads on multilingual transcription and diarization over Whisper, ElevenLabs Scribe v2, and Gemini 3.5 Transcribe, with a diarization cpWER of 18.4% and keyphrase recall on the Earnings 22 set rising to 94.6% with context versus 81.4% without. The models are offered through an EU-hosted API with self-deployment available on your own infrastructure.

Why it matters

If you are building a voice assistant or contact-center tool, this is a cheaper, EU-hosted stack with low-latency turns and an on-premises option, which matters for teams with data-residency rules. The transcription and diarization numbers come from Mistral itself, so the practical move is to run it against your current speech vendor on your own audio before committing.