Support Vector Machine Based Speech Emotion Recognition

The results obtained in this system using MFCC+LPCC with SVM arecommendable. The recognition rate ofsystem is 81.2% for IITKGP-SESC, 78.6% for EmodB and 70% for real time recorded database. The MFCCs and LPCCs corresponding to each utterance of the each emotion of databases have been computed and their fusion is used for feature extraction along with their delta and double-delta coefficients. These extracted features of training files are trained to the SVM model. Later the features of test files are given as input to SVM classifier for prediction. Then the classification of testing samples is done and the percentage of both matched and mismatched emotion is computed using confusion matrix. The performance of the real time recorded database is limited by the external factors which affect the speaker's utterances like noise in signal, environment where recording is carried out. The performance can be increased by using high quality audio devices in noise free environment. Also large number of training samples turn out to increase performance. To conclude, it can be firmly stated that despite certain limitations, this system provides appreciable efficiency and accuracy.