Clanker Me Not
Neural Network Bending in the DAW
What happens when you take a hammer to a neural network? Just as experimental musicians hacked battery-powered toy keyboards to birth circuit bending, the same subversive philosophy can be applied to neural audio. Inspired by the pioneering hardware ethos of Reed Ghazala and Nicolas Collins, and originating from early Pure Data experiments with IRCAM’s nn~, this talk explores the technical integration of neural circuit bending directly inside the DAW environment, moving past the pristine black box towards deliberate algorithmic malfunction. We will look under the hood of a real-time VST3/AU audio plugin, breaking down the C++ architecture required to intercept, modulate, and mutate internal layers, weights, and latent representations mid-inference without dropping frames or blowing up monitors. We will discuss the computational hurdles of low-latency model interaction across multiple backend runtimes and examine real-world use cases on popular architectures like autoencoders. Attendees will leave with practical engineering strategies to turn generative audio models into volatile, highly expressive musical instruments.
Leonardo Foletto
Leonardo Foletto is an independent audio software developer, live coding musician, and hardware hacker. His past work encompasses a wide spectrum of audio and music applications, from Digital Audio Workstations and audio plugins to music machine learning, automotive audio, and multimedia performance software.