Musical part separation based on perceptual hierarchy

Abstract

SUMMARYThere have been many studies of auditory scene analysis in which an attempt was made to understand external events through acoustic signals. In particular, when the target is limited to music, there have been several studies aiming at automatic music scoring. In most previous studies, however, the procedure is constantly limited to local processing. Even if the procedure continues along the time axis, it is confined to the local neighborhood on the time axis, and only limited processing performance is realized. The purpose of this paper is to improve this aspect. A method is proposed which focuses on the hierarchical structure in the perception of the acoustic stream and extracts a single note string corresponding to each part from the music. In the proposed method, phrases are formed as an intermediate step in extracting the parts. Local clues are used in forming the phrases, and global clues are used in forming the parts. By using this approach, the parts are successfully extracted without problems such as explosion of the computational complexity. In a preliminary experiment, the parts were successfully extracted with a reproduction rate of approximately 80% and a fit rate of approximately 85%.