Building a generative model for affective content of music

In the Music Information Retrieval (MIR), music emotion recognition is one of the active research topics. Many methods have been proposed for modeling the emotions from the music excerpts. Most of the methods model the emotions as a deterministic hard label which fails to handle the issue of subjectivity. Some methods which model the emotions as a distribution lack the classification [1] (Recognition in this case) accuracy. In this paper, a novel method has been proposed which models the affective content as a Gaussian probability distribution model called AEG (Acoustic Emotion Gaussian) which is generative, flexible, interpretable and transparent. AEG model incorporates two GMM's: Affective GMM used for model emotion and Acoustic GMM for modeling the audio data. In addition to the acoustic feature (MFCC) and annotations from the listeners (Valence and Arousal), additional acoustic features, Raagas (melodic modes used in Indian Classical music) and emotions obtained from the sentiment analysis of lyrics helps to optimize and increase the classification accuracy and robustness of the AEG model.