Abstract
Automatic mood detection from music has two main benefits. Firstly, having the knowledge of mood in advance can allow for possible enhancement of the music experience (such as mood-based visualizations) and secondly it makes 'query by mood' from music data-banks possible. This research is concerned with the automatic detection of mood from the electronic music genre, in particular that of drum and bass. The methodology was relatively simple, firstly sampling the music, and then giving a human pre-classification to the music (to use for training a classifier) via a point on a Thayer's model mood map. The samples then had low level signal processing features, mel frequency cepstral coefficient, psychoacoustic features and pitch image summary features extracted from them. These were then verified as useful via self organising maps and ranking via the feature selection techniques of information gain, gain ratio and symmetric uncertainty. The verified features were then used as training and testing (via cross-validation) data for a 3 layer perceptron neural network.
Two approaches at feature extraction were used due to the first approach performing poorly at self organising map based cluster analysis. The mood classification scheme was then simplified to have four moods as opposed to 25. The main difference, however between the two approaches was based around different feature extraction window duration and different features. The second approach's features were used to train the neural network and the classification performed with classification accuracy rates no less than 84 %.
Out of this research comes understanding of how one human's approximated perception can be captured and shows its use for determination of mood classifications from music.