AUTOMATIC EMOTION RECOGNITION IN SPEECH SIGNAL USING TEAGER ENERGY OPERATOR AND MFCC FEATURES

A new approach to feature extraction for the automatic emotion classification in speech signals was described and tested in this work. The method was based on the Teager energy operator combined with mel-frequency cepstral coefficients (TEO-MFCC). The proposed TEO-MFCC method was tested using speech recordings collected from the Speech Under Simulated and Actual Stress (SUSAS) database with three simulated emotions (angry, neutral and soft). The Gaussian mixture model (GMM) was used as classifier. The average classification accuracy for three emotions reached up to 73%, much higher than purely guess (33% for three emotions). Especially, the system showed good performance on the classification of emotion "angry". The correct recognition rate for emotion "angry" was 83%, while it was only 59% for emotion "neutral" and 77% for emotion "soft".