Research On Speech Emotion Recognition Based On Hybrid Algorithm Of ACON/SVM/HMM

Speech is one of the simplest and most convenient ways to express information among people. Not only semantic information, but also the speaker's emotional state can be conveyed by speech signal. Emotional information plays a significant role in the process of perceiving the external world and making some decisions. As a result, with the development of Human-Computer Interaction technology, more and more researchers pay attention to emotional information which is contained in the speech signal. Currently, speech emotion recognition, as an important part of emotional information processing has become a hotspot and received much awareness from the researchers.Speech emotion recognition can be viewed as typical pattern recognition. In this thesis, according to the basic principles of emotion recognition, several emotional features calculation and recognition algorithms are researched. Tasks which are accomplished in the thesis are listed as follows:(1) As the theoretical foundation of the subsequent content about emotional features calculation, time-domain analysis of speech waveform is firstly introduced, which contains short-time energy, short-time zero-crossing rate, short-time autocorrelation analysis and voice activity detection algorithm which is based on energy and zero-crossing rate.(2) Emotional features extraction:Two kinds of emotional features are discussed in detail in this thesis. One is instantaneous features which are based on speech signal short-time analysis, and the other is global ones. All the features are related to energy, pitch, formant and MFCC of emotional speech.(3) Emotion recognition algorithms:In this thesis, two categories of recognition algorithms are elaborated. One is All Class in One network (ACON) and Support Vector Machine (SVM), in which training and testing sets are all global features. And the other is Hidden Markov Model (HMM) method, which is trained and tested by instantaneous features.(4) Emotion recognition experiment:Firstly, according to the calculated features, the ACON, SVM and HMM are constructed as three sub-models in the training process. Secondly, the three trained emotional sub-models are combined organically by a voting strategy, and hence we get a hybrid emotional recognition model. The training samples of the proposed recognition model contain global and instantaneous emotional features. So the emotional state can be described more comprehensive.Meanwhile, the certain and uncertain recognition models are combined together, and then some shortcomings can be overcome. Finally, test speeches from the same data set are used to check the mentioned emotional recognition models. The result of the experiment shows that the speech emotion recognition algorithm based on the hybrid model can improve the performance of the emotion recognition rate effectively.