VibeVoice-ASR

VibeVoice-ASR is an open speech-to-text model for long recordings. Per the model card, it takes up to 60 minutes of continuous audio in a single pass (64K tokens) instead of cutting it into short chunks, and returns a structured transcript that says who spoke (speaker), when (timestamps) and what (content). You can pass your own hotwords, such as names and technical terms, to improve accuracy on specialized content. It supports more than 50 languages, needs no language setting and handles switching between languages inside a recording. The Hugging Face metadata lists about 8.7 billion parameters; the card does not state a VRAM requirement. The authors also published a separate edge CPU engine, VibeVoice-ASR-BitNet, and a streaming variant (see the GitHub repository). On the Open ASR Leaderboard (English short-form speech, nine test sets, checked 2026-10-20) the Transformers version has an average word error rate of 8.23%; the model is built for long recordings, so short clips do not show its main strength. Released 2026-01-21. The weights are public without any acceptance step under the MIT license, which allows commercial use. The authors also offer an official demo Space on Hugging Face, and the code is on GitHub. Date of check: 2026-10-20.