Feature selection experiments on emotional speech classification

This paper presents the experiments on feature selection for emotional speech classification. There are 152 features used in this experiment. The minimum redundancy maximum relevance (mRMR) feature selection is applied as the features selection. The experiments are constructed from two corpora; Interactive Emotional Dyadic Motion Capture (IEMOCAP) and Emotional Tagged Corpus on Lakorn (EMOLA) which are collected in English and Thai language respectively. According from the results the MFCC with ZCR present the best result of anger class (81.95% accuracy) and happiness class (69.86% accuracy). Lastly, Delta-DeltaF0 with LPREFC works best for neutral class with 67.96% meanwhile only LPREFC resulted in the highest accuracy of 80.51% in sadness class.