Matched-Filter Musical Motif Discovery

Abstract

This paper presents a system that can automatically identify all repeating parts in a song. It uniquely uses the matched filter as a self-similarity metric between the song’s segments. Unlike most musical motif discovery systems, this does not use spectral features to find similarity. Multiple experiments are used to evaluate the limits of the system’s performance under different conditions. Accuracy and precision are measured through five metrics and presented visually. We conclude that this approach is successful for short, rhythmic snippets of music with clear repetition. However, the algorithm struggles to find motifs in longer songs. Overall performance is comparable to a Shazam-based motif discovery system. This algorithm has applications in databases of songs, musical analysis, and audio thumbnailing.