A Methodology for the Computational Evaluation of Style Imitation Algorithms

We investigate Musical Metacreation algorithms by applying Music Information Retrieval techniques for comparing the output of three off-line, corpus-based style imitation algorithms. The first is Variable Order Markov Chains, a statistical model; second is the Factor Oracle, a pattern matcher; and third, MusiCOG, a novel graphical model based on perceptual processes. Our focus is on discovering which musical biases are introduced by the algorithms, that is, the characteristics of the output which are shaped directly by the formalism of the algorithms and not by the corpus itself. We describe META-MELO, a system that implements the three algorithms, along with a methodology for the quantitative analysis of algorithm output, when trained on a corpus of melodies in symbolic form. Results show that the algorithms’ output are indeed different, although none of them encompass completely the full feature-set belonging to the style of the corpus. We conclude that this methodology is promising for aiding in the informed application and development of generative algorithms for music composition problems.