On the Study of Thai Music Emotion Recognition Based on Western Music Model

Abstract

Abstract The mood of the song could be identified by tracking the listener’s emotion. The research in this area is growing significantly at the present. There are many research studies in western music, but a few in Thai music. Therefore, in this research, Thai songs were chosen because the Thai is a native language and Thai songs are quite popular in the region of research. This research is divided into 2 parts. First, Thai music was evaluated by the set of a system based on western music training settings. By using valence-arousal values, multiple linear regression, and k-nearest neighbors to represent the emotional annotations from the music. As a result, the highest f-measure of Thai music from multiple linear regression by ALL model was 41% and the f-measure of western music from multiple linear regression by No Tempo model was 51%, which was very different because ALL model in western music has lower efficiency than other models. Second, we measured the mood of 125 Thai popular songs and used valence-arousal (energy) values from Spotify API to investigate the results. In this research we used multiple linear regression (MLR) and support vector regression (SVR). Experimental results show that the multiple linear regression provides the highest accuracy of 61.29% with the precision of 65%, recall of 61%, and f-measure of 60% which is more than support vector regression.