approximations for accordion, percussion, and neural networks is a collection of five short movements featuring the sounds of two audio-generating neural networks trained on the sounds of the two performers, Louis Pino and Matti Pulkki. The neural networks (named The Accordionator and The Great Percussionist In The Sky by GPT-3) are AIs who have only ever had one sensory input: the sound of their assigned performer. They’re quirky robots trying to imitate the performers who are, in turn, trying to interact with and imitate the robots.
Movement i, “from the fog of randomness,” mimics the initial stage of training a neural network. At the beginning of the training process, the network generates statistical noise. Over hundreds of thousands of training steps, the network learns to generate sounds that more and more resemble the instruments.
Movement ii, “making friends with the robots,” takes us from rumbling machine noise to a robot dance party. The robots and performers become friends as they create a beat together, closing with a courteous, courtly dance.
Movement iii, “call + response,” gives the performers a chance to imitate the robots who have been trained to imitate them.
Movement iv, “mutual listening,” is a speculative representation of the inner/spiritual experience of the networks. These AIs aren’t conscious, though. Or are they?
The title of movement v, “educated guessing,” refers to the way networks generate audio. A network looks at a chunk of audio samples and guesses what samples should come next.
If you have questions about the technical details or want to geek out about generative AI and audio, please contact me at email@example.com.