Mistral Voxtral Transcribe 2: $0.003/min batch transcription and an open-weight realtime model
In one sentence Mistral releases Voxtral Transcribe 2: Voxtral Mini Transcribe V2 for batch transcription ($0.003/min, via API) and Voxtral Realtime, a 4B-parameter streaming model with open Apache 2.0 weights and latency configurable below 200 ms. 13 languages, Italian included.
Mistral introduces two models that turn speech into text.
The first, Voxtral Mini Transcribe V2, transcribes existing recordings: it costs $0.003 per minute via API, identifies who is speaking (diarization), adds word-level timestamps and accepts files up to 3 hours long. Mistral claims about 4% word error rate on the FLEURS benchmark and better accuracy than GPT-4o mini Transcribe, Gemini 2.5 Flash, AssemblyAI Universal and Deepgram Nova.
The second, Voxtral Realtime, transcribes live: it has 4 billion parameters, open weights under Apache 2.0 and latency configurable down to below 200 milliseconds. Mistral describes it as suitable for edge devices too.
Both support 13 languages, including Italian.
Companies
Mistral AI
Tools
Voxtral Mini Transcribe V2, Voxtral Realtime
Tags
Sources