Latent Vector Synthesis

Abstract

Generative deep learning models for sound synthesis applications have gathered interest recently that are able to generate novel sound material based on the characteristics of a given audio dataset. A subcategory of these models are variational autoencoders, which build generative latent spaces of audio where sounds are organised based on similarity. Although expressive uses of these models abound, questions around their practical applicability and aesthetic implications as part of an artistic process remain underexplored. This thesis investigates the technological and aesthetic affordances of latent audio spaces in the context of creative sound design and exploration. To this end, a sound synthesis tool in the form of a latent vector synthesizer is conceptualised and developed from a first-person research through design perspective. The prototype addresses issues around real-time playability of current machine learning models for sound generation by training a variational autoencoder on short samples of audio signals. The generated waveforms are incorporated as part of a wavetable- and vector synthesis engine that enables timbral interpolations and explorations of sonic textures. Positioned at the intersection of sonic art and audio technology the design implementation uncovers some latent potentials and affordances of new technologies for musical tasks.