Unsupervised learning of the downbeat in drum patterns

A system for the automatic determination of symbolic drum patterns along with the downbeat is presented. Fr o m an unlabeled database of over 20000 urban music songs, for each song a characteristic drum pattern of one measure length is extracted fully automatically. The 50 most frequently occurring patterns are identified. For each of the most frequently occurring patterns the downbeat is determined by investigating the cue of the drum track. An evaluation against ground truth annotations for the drum patterns is carried out, where an accuracy of 90% for the downbeat detection is achieved. Further, a listening test has been carried out, that verifies the ground truth annotations.