A Hybrid HMM And RBF Method For Speech Emotion Recognition

The new intelligent Human Computer Interaction technology has become a very active research subject in the current computer science field. The ability that the computers can understand and express emotion is one of the very important tasks. When the computer has the ability to understand the emotion, it could communicate really with humans. Speech emotion recognition is one of the key technologies about Human Computer Interaction, and it has important practical significances on the intelligent computer and the development of Human Computer Interaction system. If computers could understand the emotion and make the response, the Human Computer Interaction system will become kind and accurate.According to the deficiency of a single model, a hybrid model based on hidden markov and Radial Basis Function neural network is proposed.The main work of this thesis are as follows:(1) On the basis of the background with emotional speech banks and emotional category, an emotional speech bank including five emotional states "happiness","sadness","anger","surprise","peace" is recorded with the designated recording scripts, equipment requirements, recording personnels and etc.(2) Then according of analysing emotional characteristics, parameters extracted in this paper are as follows:energy, pitch frequency, the formant and etc.(3) The principle and method about the hidden markov model and Radial Basis Function network for speech emotion recognition are introduced detailedly. Combined with a hidden markov model with the modeling ability on dynamic sequence and Radial Basis Function network with the powerful decision-making ability,the hybrid model is builded. The hidden markov model and Radial Basis Function network are in series. The state accumulated probabilities instead of the best sequences are treated as inputs of radial basis neural network for classification and recognition.(4) The experiments by MATLAB show that the hybrid model whose average recognition rate is65.35%and5.25%more than a single hidden markov model is better. In addition,the robust testing about the hybrid model is done,and the experiments show that the performance of the hybrid model is superior to a hidden markov model.