DIRECTED AUDIO TEXTURE SYNTHESIS WITH DEEP LEARNING

Abstract

Audio textures are a group of sounds that have stable characteristics within an adequately large window of time but may be largely unstructured locally. In this thesis we develop models and techniques that allow us to synthesise a selection of audio textures while enabling the exploration and shaping of the output sound space via parameters. The utilisation of data-driven deep learning techniques improves the potential model expressivity and flexibility over existing physical and sample-based modelling approaches, both in terms of expanding the range of possible sounds and parameters by which to direct them, without having to radically change the model itself.