Learning to Play Like the Great Pianists

An application of relational instance-based learning to the complex task of expressive music performance is presented. We investigate to what extent a machine can automatically build 'expressive profiles' of famous pianists using only minimal performance information extracted from audio CD recordings by pianists and the printed score of the played music. It turns out that the machine-generated expressive performances on unseen pieces are substantially closer to the real performances of the 'trainer' pianist than those of all others. Two other interesting applications of the work are discussed: recognizing pianists from their style of playing, and automatic style replication.