Research On Speech Emotion Recognition Algorithm Based On Neural Network

In the field of computer technology,the future goal is to reduce communication barriers between people and machines.It has become a research topic with great development potential to extract the emotional features contained in the signal to determine the emotional fluctuations.Combining the extensive needs of speech emotion recognition,the speech emotion recognition algorithm based on neural network is discussed in this thesis.The specific research content is as follows:Feature extraction about four categories of speech emotions is discussed.Four categories of speech emotions are anger,fear,happiness and sadness.The four prosodic features of speech rate,short-time average energy,short-time average zero-crossing rate and pitch frequency are firstly extracted;then the sound quality feature of formant frequencies is extracted;finally,the spectrum-based correlation feature of the Mel frequency cepstrum coefficient is extracted.The algorithm of speech emotion recognition based on BP neural network is discussed.Firstly,the BP neural network is trained,in which the input is a row vector composed of different speech features,and finally the speech of different emotional categories is identified.The speech emotion recognition experiments based on BP neural network are mainly carried out in six aspects:the number of iterative training,learning rate,the number of neuron in hidden layers,the number of sample set,combination of different features,and the number of emotion category.Experimental results show that the speech emotion recognition rate is between 65.85% and 91.00% when the four different emotions are combined in pairs;the recognition rate of four different emotions in this thesis is 5.47% higher than that of algorithm based on support vector machine.The algorithm of speech emotion recognition based on convolutional neural network is discussed.Firstly,the convolutional neural network is trained,in which the feature input is a matrix composed of the emotional feature of the Mel frequency cepstrum coefficient.Finally,the speech of different emotional categories is identified.The speech emotion recognition experiments based on convolutional neural networks are implemented from two aspects: the emotion categories and the number of training set.Experimental results show that the recognition rate of the speech emotion recognition algorithm based on the convolutional neural network increases 6.50% relative to the BP neural network under four different emotions;when the four emotions are combined in pairs,the speech emotion recognition rate is between 72.00% and 97.00%,which is better than that of algorithm based on BP neural network.The experimental results of this thesis can be applied to the service industry to reduce human intervention.For example,teleconferencing,driving safety,and so on.The further study of emotional intensity can enable doctors to provide appropriate treatment for patients with mental illness.