Computational modelling of expressive music performance in jazz guitar: a machine learning approach
Sergio Giraldo · LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas) · 2016
Computational modelling of expressive music performance deals with the analysis and characterization of performance deviations from the score that a musician may introduce when playing a piece in order to add expression. Most of the work in expressive performance analysis has focused on expressive duration and energy transformations, and has been mainly conducted in the context of classical piano music. However, relatively little work has been dedicated to study expression in popular music where expressive performance involves other kinds of transformations. For instance in jazz music, ornamentation is an important part of expressive performance but is seldom indicated in the score, i.e. it is up to the interpreter to decide how to ornament a piece based on the melodic, harmonic and rhythmic contexts, as well as on his/her musical background. In this dissertation we present an investigation in the computational modelling of expressive music performance in jazz music, using the guitar as a case study. High-level features are extracted from the scores, and performance data is obtained from the corresponding audio recordings from which a set of performance actions are obtained semi automatically (including timing/energy deviations, and ornamentations). After each note is characterized by its musical context description, several machine learning techniques are explored to, on one hand, induce regression models for timing, onset and dynamics transformations, and classification models for ornamentation to render expressive performances of new pieces, and, on the other hand, learn expressive performance rules to analyse its musical meaning. Finally. we report on the relative importance of the considered features, quantitatively evaluate the accuracy of the induced models, and discuss some of the learnt expressive performance rules. Moreover, we present different approaches for semi-automatic data extraction-analysis, as well as, some applications in other research fields. The findings, methods, data extracted, and libraries developed for this work are a contribution to expressive music performance field, as well to other related fields.