Music Structure Analysis: An Exploration of and Improvements on the Distance-based Segmentation and Annotation Approach

Abstract

This thesis explores the commonly used distance-based segmentation and annotation (DSA) approach to music structure analysis. In an attempt to improve upon already existing DSA algorithms, this thesis proposes methods of combining multiple musical representations into a single fusion of features, and applies them to different approaches of distance-based segmentation and annotation. This thesis found that a fusion of a timbre feature together with a chroma feature has the most chance of increasing the quality of a DSA approach, even though the experiments in this thesis yield a statistically insignificant increase in performance. Additionally, this thesis presents numerous ways to increase performance of DSA feature fusion more. Finally, this thesis puts the DSA approach in context to other approaches of music structure analysis, such as the segmentation by annotation approach, that makes use of Deep Neural Networks. The big advantage of DSA is that it isn’t prone to over-fitting and might therefore be easier to use on multiple genres or recording types. A final discussion of the explainability of different MSA approaches is held.