Application of Change Point Detection Algorithms in Adaptable Symbolic Music Segmentation Task Using MIDI Representation

Abstract

This thesis studies the application of Change Point Detection (CPD) algorithms to the segmentation of symbolic music using MIDI representations. This study focuses on the use of two primary CPD algorithms, PELT and Binary Segmen- tation, to analyze and detect transitions within the Lakh MIDI dataset, which is known for its diversity in musical genres and styles. The main focus of the thesis is on how accurately the Binary Segmentation and PELT algorithms detect structural changes. The effectiveness of these algorithms is measured through precision, recall, and F1 score metrics, derived from manually annotated segments serving as ground truth. The adaptability of these methods across various musical structures is also evaluated to ensure their robustness and flexibility in handling different musical forms. The findings indicate that while both algorithms perform effectively, PELT shows superior adaptability and accuracy in segmenting musical structures.