Speech emotion recognition using DWT

Abstract

Emotion recognition from speech helps us in improving the effectiveness of human-machine interaction. This paper presents a method to identify suitable features in DWT domain and improve good accuracy. In this work, 7 emotions (Berlin Database) are recognized using Support Vector Machine (SVM) classifier. Entropy of Teager Energy operated Discrete Wavelet Transform (DWT) coefficients, Linear Predictive Cepstral Coefficients(LPCC), Mel Energy Spectral Dynamic Coefficients(MEDC), Zero Crossing Rate (ZCR), shimmer, spectral roll off, spectral flux, spectral centroid, pitch, short time energy and Harmonic to Noise Ratio (HNR) are considered as features. The obtained average accuracy is 82.14 % Earlier work done on emotion recognition using DWT coefficients yielded an accuracy of 63.63 % and 68.5% for 4 emotions on Berlin and Malayalam databases respectively. The proposed algorithm shows a significant increase in accuracy of about 15% to 20% for 7 emotions on Berlin database. Also, 100% efficiency has been achieved for four emotions with Simple Logistic classifier of WEKA 3.6 tool.