Wavelet Packet Entropy in Speaker-Independent Emotional State Detection from Speech Signal

In this paper، wavelet packet entropy is proposed for speaker-independent emotion detection from speech. After pre-processing، wavelet packet decomposition using wavelet type db3 at level 4 is calculated and Shannon entropy in its nodes is calculated to be used as feature. In addition، prosodic features such as first four formants، jitter or pitch deviation amplitude، and shimmer or energy variation amplitude besides MFCC features are applied to complete the feature vector. Then، Support Vector Machine (SVM) is used to classify the vectors in multi-class (all emotions) or two-class (each emotion versus normal state) format. 46 different utterances of a single sentence from Berlin Emotional Speech Dataset are selected. These are uttered by 10 speakers in sadness، happiness، fear، boredom، anger، and normal emotional state. Experimental results show that proposed features can improve emotional state detection accuracy in multi-class situation. Furthermore، adding to other features wavelet entropy coefficients increase the accuracy of two-class detection for anger، fear، and happiness.