String Quartet Classification with Monophonic Models

Polyphonic music classification remains a very challenging area in the field of music information retrieval. In this study, we explore the performance of monophonic models on single parts that are extracted from the polyphony. The presented method is specifically designed for the case of voiced polyphony, but can be extended to any type of music with multiple parts. On a dataset of 207 Haydn and Mozart string quartet movements, global feature models with standard machine learning classifiers are compared with a monophonicn-gram model for the task of composer recognition. Global features emerging from feature selection are presented, and future guidelines for the research of polyphonic music are outlined.