Intelligent Exploration of Sound Spaces Using Decision Trees and Evolutionary Approach

This paper describes Synthbee, an assistive tool for sound design which enables musicians to achieve desired sounds without managing parameters of a sound synthesizer manually. The system allows musicians to specify desired sound characteristics using attributes and explore the space of producible sounds by controlling the interactive evolutionary algorithm extended to take into account specified attributes. Using the interactive evolutionary approach, musicians can recombine and mutate patches towards a satisfactory result. While performing recombination of patches, the algorithm tries to maintain values of synthesis parameters which are relevant for achieving desired sound characteristics. Synthbee thereby enables efficient creation of novel sounds which possess characteristics described by input attributes. The method for finding and maintaining relevant synthesis parameters during an interactive exploration is our original algorithm which uniquely combines machine learning techniques with evolutionary computing. The results of the initial subjective evaluation of Synthbee showed that the users were generally satisfied with generated sounds, but also indicated some opportunities for improvement.