Melody Informatics: Computational Approaches to Understanding the Relationships Between Human Affective Reasoning and Music

Abstract

Music is a powerful and complex medium that allows people to express their emotions, while enhancing focus and creativity. It is a universal medium that can elicit strong emotion in people, regardless of their gender, age or cultural background. Music is all around us, whether it is in the sound of raindrops, birds chirping, or a popular song played as we walk along an aisle in a supermarket. Music can also significantly help us regain focus while doing a number of different tasks.

The relationship between music stimuli and humans has been of particular interest due to music's multifaceted effects on human brain and body. While music can have an anticonvulsant effect on people's bodily signals and act as a therapeutic stimulus, it can also have proconvulsant effects such as triggering epileptic seizures. It is also unclear what types of music can help to improve focus while doing other activities. Although studies have recognised the effects of music in human physiology, research has yet to systematically investigate the effects of different genres of music on human emotion, and how they correlate with their subjective and physiological responses.

The research set out in this thesis takes a human-centric computational approach to understanding how human affective (emotional) reasoning is influenced by sensory input, particularly music. Several user studies are designed in order to collect human physiological data while they interact with different stimuli. Physiological signals considered are: electrodermal activity (EDA), blood volume pulse (BVP), skin temperature (ST), pupil dilation (PD), electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). Several computational approaches, including traditional machine learning approaches with a combination of feature selection methods are proposed which can effectively identify patterns from small to medium scale physiological feature sets. A novel data visualisation approach called "Gingerbread Animation" is proposed, which allows physiological signals to be converted into images that are compatible with transfer learning methods. A novel stacked ensemble based deep learning model is also proposed to analyse large-scale physiological datasets.

In the beginning of this research, two user studies were designed to collect physiological signals from people interacting with visual stimuli. The computational models showed high efficacy in detecting people's emotional reactions. The results provided motivation to design a third user study, where these visual stimuli were combined with music stimuli. The results from the study showed decline in recognition accuracy comparing to the previous study. These three studies also gave a key insight that people's physiological response provide a stronger indicator of their emotional state, compared with their verbal statements.

Based on the outcomes of the first three user studies, three more user studies were carried out to look into people's physiological responses to music stimuli alone. Three different music genres were investigated: classical, instrumental and pop music. Results from the studies showed that human emotion has a strong correlation with different types of music, and these can be computationally identified using their physiological response.

Findings from this research could provide motivation to create advanced wearable technologies such as smartwatches or smart headphones that could provide personalised music recommendation based on an individual's physiological state. The computational approaches can be used to distinguish music based on their positive or negative effect on human mental health. The work can enhance existing music therapy techniques and lead to improvements in various medical and affective computing research. Show more