Real-time Emotion And Phoneme Recognition Based On A Two-level Model

Interaction by speech is the most natural way to communicate with computer, as the basis for it, speech recognition has always been a hot research area. Speech recognition includes speech signal pre-processing, acoustic features extraction, dimensionality reduction and statistical model based speech recognition. The thesisfocus on two major facets of speech recognition------emotion recognition and phonemerecognition. Through the improvement of existing statistical models, we finally achieve our goal: real-time recognition of emotion and phoneme.First of all, the background and content are given. Then the importance of emotion and phoneme recognition, acoustic features, statistical models and algorithms of dimensionality reduction are given.AdaBoost + C4.5, which is widely used in the areas of text classification, image recognition and so on, is introduced into speech area and improved to output two labels. Using this new model, the goal of real-time emotion and phoneme recognition is finally achieved.On the first level of emotion recognition, in order to highlight the continuity of emotion, the difference of acoustic features between two adjacent frames is introduced, in order to avoid the jitter phenomenon of emotion, a weighted voting strategy is used; On the second level of phoneme recognition, in order to avoid coarticulation, a former frame associated speech model is developed and the Isomap algorithm is also used to save time.The model described in this thesis, can be exploited in many applications such as video conferencing, games, virtual "showman", online chatting and virtual reality, etc..