Mistral AI Releases Voxtral Transcribe 2 — Ultra-Low Latency Speech Recognition Models
Mistral
Mistral AI
Mistral AI has launched Voxtral Transcribe 2, a family of speech-to-text models including Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live applications with sub-200ms latency. Voxtral Realtime is open-weights under Apache 2.0. Both models support 13 languages and offer state-of-the-art transcription quality with speaker diarization.
Mistral AI released Voxtral Transcribe 2, a next-generation speech-to-text model family. It includes Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live transcription with configurable latency down to sub-200ms. Voxtral Realtime is open-weights under Apache 2.0 license. Both models support 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, and Dutch. Voxtral Mini Transcribe V2 achieves approximately 4% word error rate on FLEURS benchmark at $0.003 per minute, outperforming competitors like GPT-4o mini Transcribe and Gemini 2.5 Flash. It also features speaker diarization, context biasing, word-level timestamps, noise robustness, and supports audio up to 3 hours. Voxtral Realtime, with a 4B parameter footprint, runs efficiently on edge devices and provides near-offline accuracy even at 480ms delay. Mistral also launched an audio playground in Mistral Studio for testing transcription with diarization and timestamps.
- Сокращения
- API = Application Programming Interface — программный интерфейс приложения
- GPU = Graphics Processing Unit — графический процессор
- LLM = Large Language Model — большая языковая модель
- TTS = Text-to-Speech — синтез речи из текста
- CRM = Customer Relationship Management — управление взаимоотношениями с клиентами
- GDPR = General Data Protection Regulation — Общий регламент по защите данных
- HIPAA = Health Insurance Portability and Accountability Act — Закон о переносимости и подотчетности медицинского страхования
Source: Mistral AI —
original