Multi-Modal Non-Prototypical Music Mood Analysis in Continuous Space: Realiability and Performances
Björn Wolfgang Schuller, Felix Johannes Weninger, Johannes Dorfner · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2011
Music Mood Classification is frequently turned into 'Music Mood Regression' by using a continuous dimensional model rather than discrete mood classes.In this paper we report on automatic analysis of performances in a mood space spanned by arousal and valence on the 2.6 k songs NTWICM corpus of popular UK chart music in full realism, i. e., by automatic web-based retrieval of lyrics and diverse acoustic features without pre-selection of prototypical cases.We discuss optimal modeling of the gold standard by introducing the evaluator weighted estimator principle, group-wise feature relevance, 'tuning' of the regressor, and compare early and late fusion strategies.In the result, correlation coefficients of .736(valence) and .601(arousal) are reached on previously unseen test data.