Emotion recognition using prosodie and spectral features of speech and Naïve Bayes Classifier

Emotion plays an important role in human communication as we can convey our feelings through it. In recent years this emotion detection from speech has become one of the major and challenging research areas in speech processing domain. Extracting suitable features from speech that influence emotions and the categorization of several emotional classes make this task more challenging. This paper proposes emotion detection using both prosodic and spectral features of speech using Naïve Bayes Classifier. Although many works consider only spectral features or prosodic features, this paper considers both. We have used pitch and Mel-Frequency Cepstral Coefficients (MFCC) as the prosodic and spectral features respectively. The classification is performed using Naïve Bayes Classifier. We have developed both gender dependent and gender independent system considering seven emotional classes. After performing the classification the accuracy of the system is tested using speech samples of popular speech database Berlin Emo-db and the results are considerably better than many existing systems.