Speech Emotion Recognition Based On Multifractal

With the rapid development of science and technology, the technology of new Human Machine Interaction (HMI) has become a very active study subject in the computer science field at present. The study of the speech emotion recognition has found important realistic value in such aspects as enhancing the intelligence and humanity of computer, developing new Human-Machine environment, and will produce good economic and social benefits.The thesis firstly introduces the study background of speech emotion recognition and the main research content, then call some key issues in the current studies of speech emotion recognition, including the overview of emotional corpus, the kinds of emotional stases, features extraction of speech emotion signals, emotional feature selection and classification algorithms. After analyzing the methods currently used by others, we firstly take the Multi-fractal theory into the speech emotional recognition, by analyzing the Multiple fractal features on the four of speech emotional (happiness,anger,sadness and neutral), and proposed Multifractal Spectrum parameters and Generalized Hurst Index as new emotional conventional parameters for speech emotion recognition. The contests are described as follows:(1) Based on the Berlin laboratory German corpus EMO-DB, We observe and analyze that speech emotions were well expressed for our analysis and experiments. Then through, we selected and defined the features(pith, resonance, energy, MFCC, etc)which are the most important in distinguishing emotions.(2) In order to overcome the inadequate of Emotional conventional linear argument at depicting different types of character sentiments,we take the Multiple fractals theory into the speech emotional recognition,By analyzing the Multiple fractal features on the different speech emotional state, and proposed Multifractal Spectrum parameters and Generalized Hurst Index. it provides a new idea for speech emotion recognition by using non-linear parameters.(3) A rough classification is taken according to the good discrimination between high intense emotion (happy and anger) and low intense emotion (sadness and neutral) of multifractal, to ensure emotions that are easily confused are grouped and to further detail the nuance among them. The rough classification creates binary intermediate nodes for SVM. Then the classification is taken on these intermediate nodes using the features of the greatest contribution, which is determinate by experience. At last, empirical characteristics based on SVM binary tree speech emotion recognition is realized ideally.