Music Analysis and Point-Set Compression

Abstract

COSIATEC, SIATECCompress and Forth’s algorithm are point-set compression algorithms developed for discovering repeated patterns in music, such as themes and motives that would be of interest to a music analyst. To investigate their effectiveness and versatility, these algorithms were evaluated on three analytical tasks that depend on the discovery of repeated patterns: classifying folk song melodies into tune families, discovering themes and sections in polyphonic music, and discovering subject and countersubject entries in fugues. Each algorithm computes a compressed encoding of a point-set representation of a musical object in the form of a list of compact patterns, each pattern being given with a set of vectors indicating its occurrences. However, the algorithms adopt different strategies in their attempts to discover encodings that maximize compression. The best-performing algorithm on the folk-song classification task was COSIATEC, with a success rate of 84%. On the other tasks, variants of SIATECCompress performed best, scoring 45% precision and 60% recall on the thematic analysis task, and 21% precision and 55% recall on the fugue analysis task.

Keywords:

• music analysis
• machine learning
• information retrieval
• folk-song analysis
• compression
• pattern discovery

Acknowledgements

The work reported in this paper was carried out as part of the EU FP7 project, ‘Learning to Create’ (Lrn2Cre8).

Notes

Department of Architecture, Design and Media Technology, Aalborg University, Rendsburggade 14, 9000 Aalborg, Denmark.

1 For definitions of chromatic pitch, morphetic pitch and other related representations of pitch and pitch-interval information, along with a discussion of how these concepts relate to other pitch representations defined in the literature, see Meredith (2006b, pp. 126–129).

2 ‘COSIATEC’ stands for ‘COmpression with SIATEC’.

3 The value of 3 for was chosen so as to be small, as the higher the value of , the more SIAR approximates to SIA. No tuning was carried out to determine an optimal value for . Collins et al. (2013) ran SIAR with .

Additional information

Funding

The project Lrn2Cre8 acknowledges the financial support of the Future and Emerging Technologies (FET) programme within the Seventh Framework Programme for Research of the European Commission, under FET [grant number 610859].