MMAU-Pro is the most comprehensive benchmark to date for evaluating audio intelligence in multimodal models. It spans speech, environmental sounds, music, and their combinations—covering 49 distinct perceptual and reasoning skills.
The dataset contains 5,305 expert-annotated question–answer pairs, with audios sourced directly from the wild. It introduces several novel challenges overlooked by… See the full description on the dataset page:
https://huggingface.co/datasets/gamma-lab-umd/MMAU-Pro.
