From algorithm to cognition: Music acoustic features and their perceptual correlates

Abstract

Music preference research seeks to explain the relationship between music listeners and their music. An important task of such research is to describe differences between types of music. Musical genres are often chosen to address this task. But they are inadequate as they require subjective interpretation by both participants and researchers, making results difficult to decipher. This thesis provides foundational work to establish Music acoustic features (MAFs). MAFs are intended to provide a reliable method of music classification and description for experimental research. First, a labelled set of 4800 musical stimuli representing six MAFs of varying levels were systematically produced. A program tool, Essentia, was then used to identify low-level audio features within the musical stimuli that correlate with MAF manipulations. The Essentia features (EFs) that best represented MAFs were identified and used to predict MAFs in 44 real-world music clips. An online study also collected ratings from participants (N = 43) for each of the 44 real-world clips. The results of MAFs predicted by EFs and MAFs rated by participants were compared for consistency. The MAF Tempo correlated strongest between predicted and rated MAFs in real-world music, followed by Dynamic, Texture, Articulation, Register, and Timbre. Based on the outlined process, MAFs were to shown to be manipulable for experimental analysis, measurable within real-world stimuli, and readily perceivable by music listeners. These three criteria firmly establish MAFs as a reliable method of music classification and description for use in experimental research. Furthermore, the process outlined here can easily be adapted to validate other potential MAFs that may exist in music. MAFs will improve future research by increasing the robustness and clarity of conclusions, and thus provide greater insight into how and why people listen to certain types of music.