Music structure analysis by subspace modeling

Automatic music structure analysis is casted as a subspace clustering problem. By assuming that the feature vectors extracted from a specific music segment are drawn from a single subspace, any sequence of such feature vectors derived from a music recording will lie in a union of as many subspaces as the music segments in the recording are. First, the sparse and the low-rank subspace clustering is tested for music structure analysis by employing three types of beat-synchronous audio feature sequences. Next, a novel computational efficient subspace clustering method is proposed, that is coined as ridge representation subspace clustering (RRSC). The performance of the aforementioned three subspace clustering methods is assessed by conducting experiments on the manually annotated Beatles benchmark dataset. The experimental results indicate that: 1) the performance of the RRSC is comparable or exceeds that of the sparse and the low-rank subspace clustering and 2) the RRSC outperforms the state-of-the-art methods proposed for music structure analysis.