Action Identity in Style Simulation Systems: Do Players Consider Machine-Generated Music As of Their Own Style?

Armen Khatchatourov, François Pachet, Victoria Rowe

The generation of musical material in a given style has been the subject of many studies with the increased sophistication of artificial intelligence models of musical style. In this paper we address a question of primary importance for artificial intelligence and music psychology: can such systems generate music that users indeed consider as corresponding to their own style? We address this question through an experiment involving both performance and recognition tasks with musically naïve school-age children. We asked 56 children to perform a free-form improvisation from which two kinds of music excerpt were created. One was a mere recording of original performances. The other was created by a software program designed to simulate the participants' style, based on their original performances. Two hours after the performance task, the children completed the recognition task in two conditions, one with the original excerpts and one with machine-generated music. Results indicate that the success rate is practically equivalent in two conditions: children tended to make correct attribution of the excerpts to themselves or to others, whether the music was human-produced or machine-generated (mean accuracy = 0.75 and = 0.71, respectively). We discuss this equivalence in accuracy for machine-generated and human produced music in the light of the literature on memory effects and action identity which addresses the recognition of one's own production.