Mistral AI has launched Voxtral Transcribe 2, a new speech-to-text model. This release includes Voxtral Mini Transcribe V2 for batch processing and Voxtral Realtime for live applications. The models boast state-of-the-art transcription quality, with features like speaker diarization, context biasing, and word-level timestamps, all available in 13 languages. The models are designed for efficiency and accuracy.

Voxtral Realtime is specifically built for low-latency applications, offering sub-200ms delay. It uses a streaming architecture for real-time transcription. The Mini Transcribe V2 model offers improvements in transcription and diarization quality. It also features enterprise-ready functions such as speaker labeling and context biasing. These tools are useful for various applications including meeting transcription and voice agents.

Voxtral Transcribe 2 also introduces an audio playground in Mistral Studio to test the new technology directly. Users can upload audio files and experiment with the features. The models' focus on efficiency and accuracy at a lower cost makes them competitive in the market. The availability of open-source models further broadens accessibility and customization options.

This release from Mistral AI is a move to compete with other transcription services. The features, language support, and emphasis on low latency position Voxtral Transcribe 2 for use in various applications, from meeting intelligence to voice assistants. The open-source aspect of Voxtral Realtime could drive further innovation and community contributions.