Some Issues on Detecting Emotions in Music

Investigating subjective values of audio data is both interesting and pleasant topic for research, gaining attention and popularity among researchers recently. We focus on automatic detection of emotions in songs/audio files, using features based on spectral contents. The data set, containing a few hundreds of music pieces, was used in experiments. The emotions are grouped into 13 or 6 classes. We compare our results with tests on human subjects. One of the main conclusions is that multi-label classification is required.