Looking Outwards – 10: Computational Music

Holly Herndon is an electronic musician who has developed the first mainstream album, called Proto, with the help of artificial intelligence. What sets her apart is that she developed a machine learning AI named Spawn to copy her own voice, and be able to sing and harmonize with her. In regards to the learning curve of the AI, Herndon states that the beginning was uninteresting – when teaching the machine, the AI extracts rules from a training canon and follows them exactly, and is unable to go outside that canon. Herndon used different programs to teach Spawn. First, she used TensorFlow, which is a visual learning program that turned soundbytes into visual spectrograms that Spawn could ‘read’. The second program Herndon used is called SampleRNN, which is used for voice recognition, and she applied it to Spawn as a way for it to predict what would naturally come next in a flow of music. The last program used was a voice model-method which involved recording herself and her partner speaking arbitrary phrases for hours of audio that translates into hundreds of megabytes of data. At the end of the process, Spawn could independently sing unique melodies that Herndon herself is not capable of replicating. I think this usage of computation for a creative practice such as this is unprecedented, as Spawn can write music that compliments Herndon’s own style, while still being something never conceived before by a human. The AI technology is not surpassing the humans in this case, but working alongside it to create something new.