Universal Classification Applied to Musical Sequences

In this paper we examine the utility of a certain type of learning techniques that are based on information-theoretic methods for music modeling. Using universal compression algorithms we apply the notion of entropy to characterize sequences in terms of their statistical source coding. This approach provides powerful methods for generation and comparison of sequences without any explicit knowledge of their statistical source. The method was applied for automatic extraction and aleatoric generation of melodies with typical motivic/melodic phrases, style mixture and style classification. The classification results show an interesting grouping that separates works of early classical period from works of a later romantic period.