Predicting Emotion from Music Audio Features Using Neural Networks

We describe our implementation of two neural networks: a static feedforward network, and an Elman network, for predicting mean valence/arousal ratings of participants for musical excerpts based on audio features. Thirteen audio features were extracted from 12 classical music excerpts (3 from each emotion quadrant). Valence/arousal ratings were collected from 45 participants for the static network, and 9 participants for the Elman network. For the Elman network, each excerpt was temporally segmented into four, sequential chunks of equal duration. Networks were trained on eight of the 12 excerpts and tested on the remaining four. The static network predicted values that closely matched mean participant ratings of valence and arousal. The Elman network did a good job of predicting the arousal trend but not the valence trend. Our study indicates that neural networks can be trained to identify statistical consistencies across audio features to predict valence/arousal values.