VibeVoice-ASR: speech-to-text model designed to handle 60-minute long-form audio (huggingface.co)

Microsoft Research has announced VibeVoice-ASR, a groundbreaking speech-to-text model capable of transcribing up to 60 minutes of continuous audio in a single pass. This innovative model generates structured transcriptions that include speaker identification (Who), timestamps (When), and spoken content (What), enhancing the usability and contextual understanding of transcriptions in various applications. Unlike traditional ASR systems that break audio into shorter segments, which often compromises coherence, VibeVoice-ASR maintains global context, improving overall accuracy and fluency. One of the standout features is its support for Customized Hotwords, allowing users to input specific terms or names to tailor the recognition process for domain-specific content. Additionally, VibeVoice-ASR can seamlessly handle over 50 languages and manage code-switching without prior language settings, making it exceptionally versatile for multilingual environments. By combining ASR, diarization, and timestamping into one model, VibeVoice-ASR represents a significant advancement for the AI/ML community, offering a tool that could transform how audio data is processed and analyzed across various sectors.