Predicting emotional response to music through a compound neural network

Abstract

The emotional effects of music have a cross-cultural component that can be explained through the tonal and non-tonal properties of musical pieces. To investigate the relationship between music and the emotions it arouses, we have built a composite neural network with the aim of predicting both the emotional categorization and the emotional valence and activation of Vieillard et al.’s (2008) musical stimuli. Our neural network uses two Adalines in the first level of the structure to predict activation and emotional valence from a minimal set of temporal and tonal properties of the stimuli (rhythm, tempo, time signature, mode, absolute tonal range and the frequency of the lowest note). In the second level, the network uses a Self-Organizing Map (SOM) network to classify the stimuli into four emotional categories (calm, happiness, fear and sadness). The results have allowed us to replicate the features of the Circumplex Model of Emotion. The percentage of explained variance obtained for activation is satisfactory and higher than in previous research for emotional valence. The percentage of music pieces correctly classified by the SOM was also very high (87%). We discuss the results in relation to competing models of music and emotion.