Improving music structure segmentation using lag-priors

Methods for music structure discovery usually process a music track by first detecting segments and then labeling them. Depending on the assumptions made on the signal content (repetition, homogeneity or novelty), different methods are used for these two steps. In this paper, we deal with the segmentation in the case of repetitive content. In this field, segments are usually identified by looking for sub-diagonals in a Self-Similarity-Matrix (SSM). In order to make this identification more robust, Goto proposed in 2003 to cumulate the values of the SSM over constant-lag and search only for segments in the SSM when this sum is large. Since the various repetitions of a segment start simultaneously in a self-similarity-matrix, Serra et al. proposed in 2012 to cumulate these simultaneous values (using a so-called structure feature) to enhance the novelty of the starting and ending time of a segment. In this work, we propose to combine both approaches by using Goto method locally as a prior to the lag-dimensions of Serra et al. structure features used to compute the novelty curve. Through a large experiment on RWC and Isophonics test-sets and using MIREX segmentation evaluation measure, we show that this simple combination allows a large improvement of the segmentation results.