Scale Degree Profiles from Audio Investigated with Machine Learning

In this paper we introduce and explore a method for extracting low dimensional features from digitized recordings of music performance: The so called constant Q scale degree profiles are 12-dimensional vectors that reflect the prominence of the 12 scale degrees in a section of a piece of music they are extracted from. Here we study the type and amount of information that is captured in those profiles when calculated from whole short pieces of piano music. The analyzed data encompass sets of preludes and fugues by Bach (WTC), Chopin (op. 28), Alkan (op. 31), Scriabin (op. 11), Shostakovich (op. 34), and Hindemith (Ludus Tonalis). In a supervised approach we investigated the ability of classifiers to recognize composers from profiles. As unsupervised methods we performed (1) a cluster analysis which resulted in one major and one minor cluster and indicated major/minor ambiguity and how clearly composers separate between major and minor, and (2) a visualization technique called Isomap which reveals in its 2-dimensional representation the degree of chromaticism of pieces apart from the major–minor duality. In summary it is astonishing how much information on a music piece is contained in the 12-dimensional profiles that can be calculated in a straight-forward manner from any digitized music recording. PURWINS ET AL. MACHINE LEARNING ANALYSIS OF MUSICAL FEATURES