Can style be learned? A machine learning approach towards ‘performing’ as famous pianists
Louis Dorard, David Hardoon and John Shawe-Taylor
In: Music, Brain & Cognition Workshop, NIPS 2007, 7-8 Dec 2007, Whistler, Canada.
In this paper a novel method for performing music in the style of famous pianists is presented. We use Kernel Canonical Correlation Analysis (KCCA), a method which looks for a common semantic representation between two views, to learn the correlation between a representation of a musical score and a representation of an artist's performance of that score. We use the performance representation based on the variations of beat level global loudness and tempo through time, as suggested by S. Dixon, W. Goebl, and G. Widmer in The Performance Worm: Real Time Visualisation of Expression based on Langner's Tempo-Loudness Animation. Therefore, the crux of the matter is the representation of the musical scores and by implication a similarity measure between relevant features that capture our prior knowledge of music. We therefore proceed to propose a novel kernel for musical scores, which is a Gaussian kernel adaptation to the distances between rhythm patterns, melodic contours and chords progressions.