Song Emotion Detection Based on Arousal-Valence from Audio and Lyrics Using Rule Based Method

Abstract

Arousal and Valence value represent of song emotions. Arousal is an emotional dimension of musically energy level, while Valence is an emotional dimension of the comfortable level of the listener. Label emotion of Thayer using Arousal and Valence dimension. This research proposed a rule base method for detecting song emotion using arousal and valence values, however many studies do not use this data. The datasets are audio and lyric features of the song structural segment chorus. Preprocessing of Audio and lyric data are uses Correlation Feature Selection (CFS) and preprocessing text. Audio feature extraction is using MIRToolbox. Stylistic and psycholinguistic are used for lyrics feature extraction. Rule based method is used to detect the emotions of the whole song by using the predictive feature of the arousal and valence values. The arousal and valence prediction values are representing withmatrices of frequencyfor audio and lyrics. From the analysis of testing data, it shows that the audio feature more represents the value of Valence while the lyrics feature more represents the Arousal value. There are seven (7) rule base models that used in this research, the best accuracy is 0.798.