Modeling Perceived Emotion With Continuous Musical Features

The relationship between musical features and perceived emotion was investigated by using continuous response methodology and time-series analysis. Sixty-seven participants responded to four pieces of Romantic music expressing different emotions. Responses were sampled once per second on a two-dimensional emotion space (happy-sad valence and aroused-sleepy). Musical feature variables of loudness, tempo, melodic contour, texture, and spectral centroid (related to perceived timbral sharpness) were coded. Musical feature variables were differenced and used as predictors in two univariate linear regression models of valence and arousal for each of the four pieces. Further adjustments were made to the models to correct for serial correlation. The models explained from 33% to 73% of variation in univariate perceived emotion. Changes in loudness and tempo were associated positively with changes in arousal, but loudness was dominant. Melodic contour varied positively with valence, though this finding was not conclusive. Texture and spectral centroid did not produce consistent predictions. This methodology facilitates a more ecologically valid investigation of emotion in music and, importantly in the present study, enabled the approximate identification of the lag between musical features and perceived emotion. Responses were made 1 to 3 s after a change in the causal musical event, with sudden changes in loudness producing response lags from zero (nearly instantaneous) to 1 s. Other findings, interactions, and ramifications of the methodology are also discussed.