Statistical Feature Selection Approach for Classification of Emotions From Speech

Abstract

Emotion plays an important role in the day to day activity of an individual and thus it’s become increasing interest and attention for many speech and language researchers. In this study author have proposed an approach of feature selection by statistical approach for the classification of emotion from a speech. The features extracted are related to energy, spectral and formant from Surrey audio visual expressed emotion dataset. Classifiers used in this study are K Nearest Neighbor, Neural Network, Random Forest, Support Vector Machine. Normality test of features has been conducted with Shapiro Wilk Test and Anderson Darling Test with 95% confidence. Features are grouped as model 1,model 2 and model 3 as per the result of normality test and principle component analysis .Results for the statistical performance of the classifier ,accuracy results into 94.99% for anger,89.11% for disgust,90% for fear,89.33% for happy,93.02% for neutral,95.24 % for sad and 90.23% for surprise .The proposed approach of feature selection gives the best results in comparison with principle component analysis except for emotion disgust and neutral.

Keywords: SAVEE speech emotion recognition Statistical test model