Classification comparison of music emotions by multiple training data sets for studying soundscape mood perception

The researches of music emotion recognition have played an important role in affective computing and content-based music information retrieval. This paper presents an approach with multiple training model to analyze the inherent emotional ingredients in the acoustic music signals, and applied to the soundscape emotion analysis. The proposed system integrated variety of emotion models including two-dimensional emotion plane and categorical taxonomy. Two sets of training data are collected, one is consisted of popular music and the other is consisted of western classical music, each set contains 192 emotion-predefined music clips respectively. Volume, onset density, mode, dissonance, and timbre are extracted from the training data to build two dimensional emotion recognition model. Gaussian mixture model (GMM) is used to demarcate the margins between four emotion states. A graphical interface with mood locus on emotion plane is established to trace the alteration of music-induced human emotion. Different sets of training data lead to the variation of boundaries among two emotion recognition models. Preliminary evaluations by Tchaikovsky 1812 Overture indicate that the emotion ingredients in the piece is consisted of 74% of Content, 19% of Depression, 5% of Anxious and 2% of Exuberant with pop song based model while the emotions composition with classical music based model is 61% of Pleasant, 24% of Solemn, 6% of Agitated and 10% of Exuberant. An extensive soundscape study is conducted by evaluating the effectiveness of mood locus variation of selected urban soundscape sets blending with music signals. Soundscape composition involves many processes, which can impact on the quality of life. Therefore, this study focuses on the challenging issue of designing an appropriate auditory atmosphere of selected soundscapes based on emotion recognition and psychoacoustic. Experimental results show that the performance of the proposed algorithms agreed well.