Voxtral Transcribe 2 (mistral.ai)

Voxtral has announced the release of Voxtral Transcribe 2, a significant upgrade in speech-to-text technology that features two new models: Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live applications. Both models offer state-of-the-art transcription quality, speaker diarization, and impressively low latency, with Realtime capable of configurable delays as low as 200 milliseconds. This capability is particularly crucial for applications requiring immediate feedback, such as voice agents and real-time communication tools. Notably, Voxtral Realtime is made available with open weights under the Apache 2.0 license, promoting widespread implementation, especially in privacy-sensitive environments. The technical advancements in Voxtral Transcribe 2 include the ability to support 13 languages and the introduction of context biasing for enhanced word accuracy. Its efficiency is demonstrated by a lower word error rate compared to competitors like GPT-4o and Deepgram Nova, while processing audio three times faster than ElevenLabs’ Scribe v2, all at a competitive price point. The models also feature capabilities crucial for enterprise applications, such as precise word-level timestamps and noise robustness, setting a new standard for transcription and diarization across various industries, from media to compliance and call centers.