Relating Perceptual and Feature Space Invariances in Music Emotion Recognition

It is natural for people to organize music in terms of its emo- tional associations, but while this task is a natural process for humans, quantifying it empirically proves to be a very dicult task. Consequently, no particular acoustic feature has emerged as the optimal representation for musical emotion recognition. Due to the subjective nature of emotion, determining how informative an acoustic feature domain is requires eval- uation by human subjects. In this work, we seek to perceptually evaluate two of the most commonly used features in music information retrieval: mel-frequency cepstral coecients and the chromagram. Furthermore, to identify emotion-informative feature domains, we seek to identify what musical features are most variant or invariant to changes in musical qual- ities. This information could also potentially be used to inform methods that seek to learn acoustic representations that are specifically optimized for prediction of emotion.