Emotional Speech Processing In The Human-machine Communication

With the development of information science and computer technology, Human-computer communication becomes the key technology for computer system to be personate and intelligent. Speech is one of the most natural and effective modal for human-computer communication. Speech communication doesn't depend on any display devices and the output terminals are cheap, common and easy to take. Thus, speech interface has potential and bright prospect in the field of multi-modal human-computer communication, which is the developing tendency of GUI (Graphic User Interface). Emotion information is an important part of speech signals and is ignored totally and even deleted directly in traditional speech processing, which becomes an obstacle on the way of speech research to application field. The ability to process affective speech is an indication of intelligent human-computer communication technology. This is significant to improve the results of speech synthesis or recognition, and raise robustness of speech processing or speech communication systems.In the beginning of this paper, the research background and history of affective speech processing are reviewed. And then some usual methods of classifying emotion spaces and primary algorithms of affective speech processing are introduced. Through the comparison between the algorithms of emotion recognition and affective speech synthesis, the most proper ones are selected and improved. Time parameters, pitch parameters and energy parameters are used as prosodic parameters in the statistical analysis, which are extracted from a multilingual speech corpus consisting of English, Chinese and Japanese affective speech samples. The characteristics of these prosodic parameters are compared and concluded, and based on which the separability and recognition of emotion information in speech signals are researched.Emotion information recognition and affective speech synthesis in human-computer communication based on speech interaction are the most important two aspects in the research of this paper. In emotion information recognition experiments, the statistical of prosodic parameters results show that there are obvious diversities and distribution under different emotion states, and language factor does not affect the acoustic correlates of affective speech. In our research we make effort to...