SEMANTIC SEGMENTATION OF MUSIC AUDIO CONTENTS
Bee Suan Ong, Perfecto Herrera · 2005
ABSTRACT This paper proposes a novel approach to detect structural changes in music audio signals and provides a way to separate the different “sections” of a piece, i.e. “intro”, “verse” or “chorus”. Herein, we divide the segment boundaries detection task into a two-phase process with each having different functionalities. In order to obtain appropriate structural boundaries, we propose a combination of low-level descriptors to be extracted from music audio signals. A database of 54 audio files (The Beatles’ songs) is used for evaluation on a mainstream popular music collection. The experiment results show that our approach has achieved 71% of accuracy and 79% of reliability in identifying structural boundaries in music audio signals. These measures indicate that the performance of our method improves the results reported in the still scarce literature that includes quantitative analyses. 1. INTRODUCTION Music structure varies widely from composer to composer and from piece to piece. Transformation, repetition, elaboration and simplification of music materials help to create the unique identity of music. Hence, it is believed that structural description provides a powerful way of interacting with audio content (i.e. browsing, summarizing, retrieving and identifying). Seeing this uniqueness of music structure, it is interesting to ask the question: is it possible to detect non-trivial/significant structural changes (i.e. intro->verse, verse->chorus, chorus->bridge, etc.) in music audio signals? This paper presents a novel, two-phased approach to this problem based on audio content analysis and similarity computation. In order to obtain appropriate musical content descriptions to detect structural changes, we propose a combination set of low-level descriptors to be extracted from music audio signals. In this paper, we address the problem of finding acceptable structural boundaries, without prior knowledge about musical structure. There is a second related problem consisting on assigning labels to the found segments. This will be reported in the first coming publication.