Brave: Designing an Embedded Network-Bending Instrument, Manifesting Output Diversity in Neural Audio Systems

As neural audio synthesis becomes more widely adopted there is a growing risk that its limitations could impact the content, quality and diversity of mu- sic. Some musicians, artists, and researchers perceive an increased risk of cultural homogenisation and qualitative degeneration due to poor quality training data and parameterisation. This work seeks to explore new methods for addressing these challenges by contributing to the developing field of ”network-bending”. Network-bending employs direct manipulation of internal ML architectures to enable active divergence from the training corpus, increasing the statistical variability and capability of model outputs. We present “Brave”: an embedded, network-bending hardware instrument, which can provide a novel blueprint for embedding a network- bending system on a stand-alone system. Through a process of iterative musician-led feedback, drawing on Proof-of-Concept Media and Arts Technology approaches, this work seeks to stimulate further interest in network-bending frameworks applied in the field of AI-driven sound synthesis.