Developing methods for predicting affect in algorithmic composition

Abstract

Affective Algorithmic Composition (AAC) is a field that focuses on the algorithmic generation of music specifically to affect its audience in a targeted way. This thesis presents a novel method for developing AAC systems based on collecting both perceived and induced affect data from human participants using multiple biosensor and surevey approaches, and modelling the resulting data in a predictive function based on a neural network. This in turn is used to drive the musical algorithm to generate music that can invoke any specified affective target. These various approaches to affect measurement can be assessed and compared by their respective predictive error when used to train a neural network, providing an assessment tool for further refinement and development. A pilot study of this method is also presented, The Affective Algorithmic Composer (AACr). AACr‟s predictive functions are trained using multiple forms of affect data collected from a group of participants, and can generate original music to invoke specific emotional states, physiological states, perceived content, and themes. Several generated compositions are included to demonstrate the abilities of the AACr to invoke affective states defined manually or directly taken from the user via biosensors. The thesis concludes by reflecting on the method‟s strengths, areas for further development, and methods that could be used to determine the success of future AAC systems.