Emoti̇on recogni̇ti̇on from speech si̇gnal

Conversation signals are considered as one of the fastest and the natural communication methods among people. This case raised the importance of identifying emotions through conversation signals for researchers in order to make human and machine communication quicker and more efficient. In this thesis study, emotion classes like Angry, Neutral, Happy and Sad also a data base, called as EmoSTAR, which consist of totally 393 data and contains Turkish-English examples, are mentioned. Having two different language examples is sufficient in terms of showing emotions independent from pronunciation and language. Using this data base, it is investigated with different features adding Mel Frequency Cepstral Coefficients (Mfcc), in addition, zeroth Mfcc, energy and first-second derivatives from each speech signal. Furthermore, while Mfcc is extracting, the length of frame and scroll-time were changed in order to study the effect of it on the results. Also, in this thesis study, analyses are made by using Hu Moments and Linear Prediction Coefficient (LPC) features. Obtained features are evaluated using Support Vector Machines (SVM), K Nearest Neighbor (k-NN) classifier and cross-validation method and success rate was obtained as %98,7. Also, in this study EmoDB was used as a test set and verification between different database was performed. The final phase of this study, dimension reduction process has been done by principal componenet analysis and thus it is seen that good results has obtained in terms of processing time and succes rate.