Continuous Affect Responses to a Large Diverse Set of Unfamiliar Music: Bayesian Time-Series and Cluster Analyses

Sixty-nine participants made continuous response judgments of perceived arousal and valence while listening to 30-s extracts of 100 unfamiliar pieces within a novel recommender system. Our purpose was to take advantage of the relatively large number of participants and pieces studied (compared with prior work on time series analysis [TSA] of continuous affective responses to music) to test the generality of the deductions made from prior evidence in smaller studies. Accordingly, we expected highly autoregressive responses, and hypothesized that acoustic intensity fluctuation would be a major (population-wide, fixed-effect) predictor of arousal and to a lesser degree valence responses in multivariate models, operating in conjunction with broad features of the spectral profile of the music (such as spectral flatness and centroid). We expected considerable interparticipant variability and were interested in the degree to which there was also interpiece variability (both being features of group-specific, random effects). The basic expectations were supported, and we also demonstrated, partly by means of a novel use of cluster analysis, that the observed group variabilities in response features could not be readily attributed to indices of a range of participant-preference and piece-feature characteristics, determined prior to the experiments. A Bayesian analysis was developed, possibly for the first time in the TSA of continuous affect responses to music, since it provides a more comprehensive description of the parameters of the analysis (not only their best estimate point values) and their effects, and a comprehensive range of tests for meaningful effects.