Visualization of song collections for understanding of the relationship between metadata and features

Abstract

When performing automatic classification and recommendation for music, there exists a problem that it is not clear which metadata and features of songs should be selected, and it is difficult how to discuss the usefulness of each element.Therefore, we tackle the visualization for songs and its metadata to examine the effective factors (metadata, acoustic features, machine learning methods and visualization methods) for music classification tasks and to discover new relationships between acoustic features and metadata.In this study, we discuss the result of the visualization about acoustic features and metadata distributions calculated from any group of songs using music signal processing tools and machine learning methods.And then, we show some experimental results using metadata like release date and composer name and acoustic features calculated from tempo, average value of the sound volume, and machine learning methods.