If Turing Played Piano With an Artificial Partner

Music is an inherently social activity that allows people to share experiences and feel connected with one another. There has been progress in neural network architectures that implement large-language-like generative models producing realistic musical scores, but less progress has been made to enable social experiences for a human playing interactively along with an artificial partner. Playing music socially sometimes works better without a score: Each participant must complement the ideas of other musicians. We investigated whether a generative model trained for passive production of musical scores, not for interaction, could enable social interaction with high quantitative measures of experience. The model, a variational autoencoder pretrained on a very large corpus of digitized and quantized piano scores, was adapted for a timed call-and-response task with a human partner. Human participants played piano with another human and separately with artificial partners parameterized for different time spans, tendencies for variation, and imitation of the human. Improvisation was not required from the humans, but it was encouraged. After each trial, participants rated the performance quality and their experience of self-other integration. Overall, the artificial partners failed at enabling convincing improvisation but were seen as adequate for basic practice. Performance differences among configurations of the artificial partner suggested paths for the future evolution of such partners. The ones with the simplest architecture and highest similarity were rated highest and were not statistically different from the human partners on realism, ease to interact with, and self-other integration. We discuss what principles may be needed, including open-ended evolution in Artificial Life, to arrive at artificial partners that enable genuinely interactive experience.