Using Perceptual Timbre Descriptors to Browse Large Audio Collections

We investigate the use of perceptually-modelled descriptors of timbre for browsing large collections of user-contributed sound. Such descriptors are already available in the sound-sharing site Freesound through their public API but have been largely unexplored for regular use. A large-scale analysis of the statistical behavious of these descriptors is presented. We believe such analysis has not been performed before in MIR outside of the context of model-driven feature extraction. Such analysis is then used to inform the design of a search results-filtering interface, based on range sliders, that makes use of these timbral descriptors. We end the study with an evaluation of such interface on actual Freesound users through the use of online surveying. We believe our proposed browsing interface using perceptually-modelled descriptors can provide novel and useful ways of navigating large sound collections