Evaluating Low-Dimensional Latent Representations as a Creative Interface for Digital Synthesizers

Synthesizers are musical instruments that typically expose a large set of user-adjustable parameters that shape the sound of the instrument. In this paper, we examine how leveraging low-dimensional latent representations of synthesizer patches enables musicians to generate new patches via exploration of a generative model's latent space. We evaluate two different latent representations, Latent Coordinates and Timbral Representation, through a mixed-methods user study (n=18). In our study, a mix of novice and experienced musicians engage with both sound matching and sound discovery tasks. Our qualitative results highlight a number of key themes regarding the suitability of our technique with each being supported through related quantitative results. Finally, these results indicate several ways in which this type of support tool may be used in a musician's existing creative workflow as well as provide a context for discussing the benefits of continuous interactions for generative systems as it relates to creative activities.