The Impact of Pre-processing and Feature Selection Methods for Speech Emotion Recognition

Speech emotion recognition uses features obtained from digital signal processing and digitized sound signal. All of the features extracted from the speech can be handled under one dimension as well as grouped in terms of dimensional or structure. In this study, the effects of feature selection and preprocessing methods on emotion detection were investigated. For this purpose, EMO-DB data set and three different classifiers are used. According to the results obtained, the highest success was achieved with 90.3% with multi-layer perceptron and high-pass filter. Spectral features provide higher success than prosodic features. In addition, females compared to males and individuals in 20-29 age interval compared to individuals in 30-35 age interval reflect their emotions more to their voices. Among the filtering methods obtained in the study, high-pass filtering increased the success of classifier whereas low-pass filtering, band-pass filtering and noise reduction reduced it