Automatic speech emotion recognition based on hybrid features with ANN, LDA and K_NN classifiers

Despite many efforts in Speech Emotion Recognition, there is still a big gap between natural human feelings and computer perception. In this article, the recognition of the speaker's emotions in Persian and German has been examined. For this purpose, Persian emotional speech utterances have been expressed, including 748 sentences with seven feelings of Neutral, Disgust, Fear, Anger, Sadness, Boredom and Happiness. German emotional speech utterances consist of 536 sentences created by professional actors in a laboratory environment, 16 of which with seven different feelings of Happiness, hatred, naturalness, fear, Sadness, Anger, and fatigue. After extracting widely used properties such as MFCC Mel Frequency Cepstral Coefficients and its derivatives, local frequency perturbation coefficient (Jitter), and local perturbation coefficient (Shimmer), various features of this database are extracted separately because of the vast number of options. Reducing feature space is required before applying the principal component classification (PCA) algorithm. Also, three classifications of Artificial neural network (ANN), Linear Discriminant Analysis (LDA), and K_Nearest Neighbor (K_NN) have been used to classify emotions. For the German database, the top results were obtained by fusing the MFCC + Shimmer properties and LDA classification with a precision detection of 91.26% and a runtime execution of 0.43 s, and the best results for the Persian database were obtained by fusing the Jitter + Shimmer properties and K_NN classification with a precision detection of 91.5% and a runtime execution of 0.65 s. The results show that the ability to distinguish attribute vectors is quite different for each emotional state. Expression of emotions and their effect on speech differ in Persian and German.