Databionic visualization of music collections according to perceptual distance

We describe the MusicMiner system for organizing large collections of music with databionic mining techniques. Low level audio features are extracted from the raw audio data on short time windows during which the sound is assumed to be stationary. Static and temporal statistics were consistently and systematically used for aggregation of low level features to form high level features. A supervised feature selection targeted to model perceptual distance between different sounding music lead to a small set of non-redundant sound features. Clustering and visualization based on these feature vectors can discover emergent structures in collections of music. Visualization based on Emergent Self-Organizing Maps in particular enables the unsupervised discovery of timbrally consistent clusters that may or may not correspond to musical genres and artists. We demonstrate the visualizations capabilities of the U-Map, displaying local sound differences based on the new audio features. An intuitive browsing of large music collections is offered based on the paradigm of topographic maps. The user can navigate the sound space and interact with the maps to play music or show the context of a song