ℓ1-GRAPH BASED MUSIC STRUCTURE ANALYSIS

Anunsupervisedapproachforautomaticmusicstructureanalysis is proposed resorting to the following assumption: If the feature vectors extracted from a specific music segment are drawn from a single subspace, then the sequence of feature vectors extracted from a music recording will lie in a union of as many subspaces as the music segments in this recording are. It is well known that each feature vector stemming from a union of independent linear subspaces admits a sparse representation with respect to a dictionary formed by all other feature vectors with nonzero coefficients associated only to feature vectors that stem from its own subspace. Such sparse representation reveals the relationships among the feature vectors and it is used to construct a similarity graph, the so-called l1-graph. Accordingly, the segmentation of audio features is obtained by applying spectral clustering to the l1-graph. The performance of the just described approach is assessed by conducting experiments on the PopMusic and the UPF Beatles benchmark datasets. Promising results are reported.