Algorithms can Mimic Human Piano Performance: The Deep Blues of Music

Abstract

Can a computer play a music score, e.g. via a Disklavier, in a way that cannot be distinguished from a human performance of the same music? One hundred and seventy-two participants with a wide range of music playing backgrounds rated sound recordings of 7 performances of piano music by Kuhlau, one played by a human, and six generated by algorithms, including a ‘mechanical’ and an ‘unmusical’ rendering. Participants rated the extent to which each performance was by a human and explained their answers. The mechanical performance had the lowest mean rating, but the human performance was rated as statistically identical to the other stimuli. There were no differences between ratings made by classical piano experts and lay listeners, but despite this, the musicians were more confident with their ratings. Qualitative analysis revealed five broad themes that contribute to judging whether a piece appears to be human. The themes were labelled (in descending order of frequency) intuitive, expressive, imperfections, halo (global preference) and empathy. This paper presents new evidence systematically demonstrating that algorithm generated performances of piano music can be indistinguishable from human performances, suggesting some parallels with the 1990s victory of the Deep Blue computer of the world champion (human) chess player.

Keywords:

• automated piano performance
• music judgement
• music competitions
• musical Turing test
• expert-novice
• artificial intelligence