Mining Unlabeled Electronic Music Databases through 3D Interactive Visualization of Latent Component Relationships

Abstract

We present an interactive content-based MIR environment specifically designed to aid in the exploration of databases of experimental electronic music, particularly in cases where little or no metadata exist. In recent years, several rare archives of early experimental electronic music have become available. The Daphne Oram Collection contains one such archive, consisting of approximately 120 hours of 1/4 inch tape recordings and representing a period dating from circa 1957. This collection is recognized as an important musicological resource, representing aspects of the evolution of electronic music practices, including early tape editing methods, experimental synthesis techniques and composition. However, it is extremely challenging to derive meaningful information from this dataset, primarily for three reasons. First, the dataset is very large. Second, there is limited metadata some titles, track lists, and occasional handwritten notes exist, but where this is true, the reliability of the annotations are unknown. Finally, and most significantly, as this is a collection of early experimental electronic music, the sonic characteristics of the material are often not consistent with traditional musical information. In other words, there is no score, no known instrumentation, and often no recognizable acoustic source. We present a method for the construction of a frequency component dictionary derived from the collection via Probabilistic Latent Component Analysis (PLCA), and demonstrate how an interactive 3D visualization of the relationships between the PLCA-derived dictionary and the archive is facilitating researcher’s understanding of the data.