Prediction of Multidimensional Emotional Ratings in Music from Audio Using Multivariate Regression Models

Content-based prediction of musical emotions and moods has a large number of exciting applications in Music Information Retrieval. However, what should be predicted, and precisely how, remain a challenge in the field. We provide an empirical comparison of two common paradigms of emotion representation in music, opposing a multidimensional space to a set of basic emotions. New groundtruth data consisting of film soundtracks was used to assess the compatibility of these models. The findings suggest that the two are highly compatible and a quantitative mapping between the two is provided. Next we propose a model predicting perceived emotions based on a set of features extracted from the audio. The feature selection and transformation is given special emphasis and three separate data reduction techniques are compared (stepwise regression, principal component analysis, and partial least squares regression). Best linear models consisting of 25 predictors from the data reduction process were able to account for between 58 and 85% of the variance. In general, partial least squares models performed the best and the data transformation has a significant role in building linear models.