Compression-based geometric pattern discovery in music

Abstract

The purpose of musical analysis is to find the best possible explanations for musical objects, where such objects may range from single chords or phrases to entire musical corpora. Kolmogorov complexity theory suggests that the best possible explanation for an object is represented by the shortest possible description of it. Two compression algorithms, COSIATEC and SIATECCOMPRESS, are described that take point-set representations of musical objects as input and generate compressed encodings of these point sets as output. The algorithms were evaluated on a task in which 360 folk songs were classified into tune families using normalized compression distance, a 1-nn classifier and leave-one-out cross-validation. COSIATEC achieved a success rate of 84% on this task, compared with a success rate of 13% for a general-purpose compressor. Variants of the algorithms incorporating modifications that have been suggested in the literature were also run on the task and the results were compared.