Voxtral Transcribe 2
2026-02-04 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Voxtral Transcribe 2, two speech transcription models: Voxtral Mini Transcribe V2 in batch mode and Voxtral Realtime in real time, the latter under the Apache 2.0 license.
What changes
- Two models: Voxtral Mini Transcribe V2 (batch transcription) and Voxtral Realtime (live applications, latency below 200 ms possible).
- Speaker diarization with precise timestamps, word-level timestamps and context biasing (up to 100 words or phrases).
- Noise robustness and audio up to 3 hours per request.
- 13 languages supported.
### Specifications - Languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, Dutch. - License: Apache 2.0 for Voxtral Realtime (weights on Hugging Face); Voxtral Mini Transcribe V2 available via API under the Premier license (proprietary, according to the Mistral documentation). - Availability: Mistral API, Mistral Studio (audio playground), Le Chat. - Context, maximum output and knowledge cutoff: not applicable or not disclosed. ### Pricing | Model | Price | |---|---| | Voxtral Mini Transcribe V2 | 0.003 USD per minute | | Voxtral Realtime | 0.006 USD per minute | ### Benchmarks | Measure | Result | |---|---| | Word error rate on FLEURS (Mini Transcribe V2) | about 4% | Scores published by the vendor. Mistral also claims to outperform GPT-4o mini Transcribe, Gemini 2.5 Flash and Assembly Universal in accuracy, without detailed figures included here. ### Key takeaways An open transcription model (Realtime) deployable at the edge keeps audio data in-house, while the batch version targets high-volume, low-cost use cases.
Links
Tags : LLM, Mistral, Voxtral, Audio, Transcription, Open Source