Latent Novelification: Expanding Generative Expressivity Through Objective-Defined Transformation

Musicians are using neural audio models to synthesize new sounds, but the sound space accessible through such models is often constrained to essentially interpolated approximations of existing training data, limiting the true sonic novelty musicians can elicit from generative models. To address this problem, we present Latent Novelification (LN), a machine learning method which learns to transform vectors in any generative model's latent space into new, contrasting vectors according to an objective function which we define. The LN objective enforces that these contrasting vectors explicitly diverge from existing training data while remaining semantically meaningful to the generative model, facilitating generative novelty beyond simple interpolation. LN supports divergent creativity by helping musicians generate new and novel sounds, integrating into any neural audio model interface to controllably and interactively expand accessible sound space. This work contributes a new algorithmic method---implemented in open-source software---which musicians working with neural audio synthesis can use to synthesize new sounds to incorporate into their art.