Statistical Music Structure Analysis Based on a Homogeneity-, Repetitiveness-, and Regularity-Aware Hierarchical Hidden Semi-Markov Model
Go Shibata, Ryo Nishikimi, Eita Nakamura, Kazuyoshi Yoshii · Zenodo (CERN European Organization for Nuclear Research) · 2019
This paper describes a music structure analysis method that splits music audio signals into meaningful segments such as musical sections and clusters them. In this task, how to model the four fundamental aspects of musical sections, i.e., homogeneity, repetitiveness, novelty, and regularity, in a unified way is still an open problem. Here we propose a solid statistical approach based on a homogeneity-, repetitiveness-, and regularity-aware hierarchical hidden semi-Markov model. The higher-level semi-Markov chain represents a sequence of sections that tend to have regularly spaced boundaries. The timbral features in each section are assumed to follow emission distributions that are homogeneous over time. The lower-level left-to-right Markov chain in each section represents a chord sequence whose sequential order is constrained to be a repetition of a chord sequence in another section of the same cluster. The whole model can be trained unsupervisedly based on Bayesian sparse learning where unnecessary sections automatically degenerate. The proposed method outperformed representative methods in segmentation and clustering accuracies with estimated sections having similar statistical properties as the ground truth data.