Voxtral transcribes at the speed of sound.

Mistral is making a real push into the voice AI stack, and its latest release, Voxtral Transcribe 2, is aimed squarely at the gap between cheap-but-slow transcription and expensive-but-fast. The lineup splits into two: Voxtral Mini Transcribe V2 for batch jobs like meeting recordings, and Voxtral Realtime, a streaming model built for live audio. Mistral says Realtime can hit sub-200ms latency, which is fast enough to power voice agents that don't feel like they're waiting on a satellite delay.

What makes Realtime interesting isn't just the speed, it's the architecture. Instead of chopping audio into chunks and running an offline model on each piece, like most competitors do, it transcribes audio continuously as it streams in. At 480ms delay, Mistral claims word error rates within 1-2% of near-offline accuracy, and at 2.4 seconds delay, it matches its own batch model, Mini Transcribe V2, making it viable for subtitling too. It's a 4B parameter model, small enough to run on edge devices, and it's shipping under Apache 2.0 on Hugging Face, so anyone can self-host it for privacy-sensitive use cases without leaning on Mistral's servers.

The batch model, Mini Transcribe V2, is where Mistral is making its boldest pricing claim. At roughly 4% word error rate on the FLEURS benchmark and $0.003 per minute, the company says it beats GPT-4o mini Transcribe, Gemini 2.5 Flash, AssemblyAI's Universal, and Deepgram's Nova on accuracy, while running about three times faster than ElevenLabs' Scribe v2 at a fifth of the cost. Both models now support 13 languages, including Mandarin, Hindi, Arabic, and Japanese, and Mini Transcribe adds speaker diarization, word-level timestamps, and something called context biasing, where you feed the model up to 100 words or phrases so it stops mangling names and jargon. That last feature is optimized for English for now, with other languages still experimental.

Mistral also rolled out an audio playground inside Mistral Studio, letting anyone upload files and toggle diarization, timestamps, and bias terms without touching an API. It's a small addition, but it signals that Mistral wants developers testing this stuff casually, not just integrating it blind. The target use cases read like a checklist of enterprise pain points: meeting transcription, contact center automation, live subtitling, compliance audit trails, all pitched as GDPR and HIPAA-friendly through private deployment options.