Research Of Emotion Recognition Based On Combined Speech Feature

Speech is one of the important ways for human's communication and it is a convenience and simple way to transmit information. The speech signal contains not only the expressed speech meaning, but also the speaker's emotion information which always be ignored by the traditionally speech processing. But the emotion information plays a very important role in the speech communication. So in recent years, emotion recognition has become a hotspot.This thesis researches two kinds of recognition models, one of them is ACON (All Class in One neural Network) network based on BP (Error Back Propagation) algorithm, the other is DTW (Dynamic Time Warping) algorithm based on template matching. A speech emotion recognition system based on BP neural network algorithm is designed and implemented. The main work of this thesis is as follows:(1) Basic knowledge of speech signal processing related to speech emotion recognition is introduced.(2) The analysis and extraction of emotion feature parameters is discussed in detail. An accurate emotion feature extracting is the foundation for computer to identify the emotion state correctly. According the difference of different emotional pronunciation mechanism, the thesis selects several different types of parameters: amplitude energy, pitch frequency, formant, LPCC (Linear Predictive Cepstral Coefficient) and MFCC (Mel Frequency Cepstral Coefficient, MFCC). After refining the parameters based above on, their derivative features are given.(3) Two kinds of emotion recognition algorithm are researched:BP algorithm and DTW algorithm. The simulation comparison experiments are designed and completed by MATLAB 7.0, the results show that traditional DTW algorithm is not so stable as multi-template DTW algorithm, either not able to get higher recognition rate. However, the BP algorithm is more efficient and reliable than the multi-template DTW algorithm.(4) An online speech emotion recognition system based on BP algorithm is implemented by VC++6.0 preliminarily. The system can identify the speaker's emotion state which depends on the trained emotion recognition network and the current input emotion speech signal. The different cartoon countenances will be shown as the emotion recognition result.