Communication Technology for Voice Emotion Detection Based on Hybrid Deep Learning Method

Abstract

Voice is an important way to express emotions, while the emotive information contained in the voice under natural and performance state is not identical. To explore the differences between vocal emotion tests under natural and performance state and to improve the accuracy of intelligent vocal emotion test system, a kind of vocal emotion test model based on the feature representation of deep learning CNN(Convolution Neural Network) is put forward. With Lenet－5 model as the foundation, a convolutional layer and a pooling layer are added to the above convolution model, meanwhile the two-dimensional convolution kernel is changed into a one-dimensional one, which is conveyed to the above model after pre-processing the one-dimensional features, representation of the features is completed, and finally sentiment classification is realized through a SoftMax classifier. It is indicated in the experimental results that the emotion recognition rate under natural state is obviously higher than that under performance state, meanwhile misclassification rate of anger and sadness under both states is also found.