Recognition of Musically Similar Polyphonic Music

Abstract

When are two pieces of music similar? Others have tackled this problem either by considering the acoustic signals of musical performances, or by looking at features of a symbolic rendition of the piece, either as MIDI data or as some direct representation of the music score. This paper presents a new approach to assessing the similarity of polymorphic music segments by combining a feature-driven clustering approach with one that measures the contrapuntal similarity of the segments. On a composer classification task, our techniques achieved almost 80% accuracy when applied to a large database of short music segments from four classical composers. This is a significant improvement to other work on composer classification based on melodic themes