Segmental pattern discovery in music

In this paper we describe a new method for discovering recurrent patterns in a corpus of segmented melodies. Elements of patterns in this scheme do not represent individual notes but rather represent melodic segments that are sequences of notes. A new knowledge representation for segmental patterns is designed, and a pattern discovery algorithm based on suffix trees is used to discover segmental patterns in large corpora. The method is applied to a large collection of melodies, including Nova Scotia folk songs, Bach chorale melodies, and sections from the Essen folk song database. Patterns are ranked using a statistical significance method that integrates pattern self-overlap, length, and frequency in a corpus into a single measure. A musical interpretation of some of the statistically significant discovered patterns is presented.