Automatic Audio Segmentation aims at extracting information on a song's structure, i.e., segment boundaries, musical form and
semantic labels like verse, chorus, bridge etc. This information can be used to create representative song excerpts or summaries, to facilitate browsing in large music collections or to improve results of subsequent music processing applications like, e.g., query by humming.
This paper features algorithms that extract both segment boundaries and recurrent structures of popular songs. Special attention has been paid to the evaluation setup: We employ the largest corpus that has been used so far in this field, discuss why comparing two song segmentations is inherently delicate and propose a
exible XML format that can describe hierarchical segmentations to promote a common basis that makes future
results more comparable.