MusicNet Music Dataset

MusicNet is a large music dataset used to monitor and evaluate machine learning methods for music research. The dataset consists of 330 copyright-free classical music recordings and over 1 million annotation labels. The exact time when each note appears on each recording, the instrument that plays the note, and the position of the note in the rhythmic structure of the work are noted. These labels were evaluated and verified by musicians, and the error rate of the labels was only 4%.