Multi-layer SVM Speech Emotion Recognition Based On Genetic Optimization

Language is the main medium of human communication.It not only contains abundant semantic information,but also carries a wealth of emotional information.It is of great significance to study how to make the computer identify the speaker's emotional state from the speech signal,so as to realize the natural human computer interaction.Aiming at the problem of slow recognition speed and low recognition accuracy in speech emotion recognition,in this paper,we proposes a scheme of classifying multi class emotion by constructing a multi-layer SVM with binary tree structure.And we use the genetic algorithm to reduce the dimension of the feature,so as to further improve the recognition performance of the speech emotion recognition system.In this paper,we first extract the speech energy,pitch period,formant and MFCC after the speech signal is processed that incluing pre-emphasi,windowing and framing,endpoint detection.Then,the hierarchical SVM with binary tree structure is constructed for multi class emotion classification.This model first divides the easily distinguishable emotions,and then classifies the easily confused emotions.The classification of multi class emotion is realized by layer by layer.The experiment was carried out on the Berlin emotional corpus containing 7 emotions.The results show that the model has the advantages of high classification efficiency of SVM,and also has the characteristics of high calculation efficiency of binary tree structure.Different features have different ability to distinguish emotions.And it is easy to lead to overfitting in modeling when the feature dimension is too high,which leads to long modeling time and low recognition accuracy.Therefore,we can optimize the extracted emotional features and then train the classification model to further optimize the multi-level SVM with binary tree structure.In this paper,we use the genetic algorithm to reduce the feature dimension,which is to select the key features from the extracted features.This method is an adaptive global optimal search method,and does not change the value of the selected feature,therefore,a better model can be constructed.The experiment is also carried out on the Berlin emotional corpus.The results show that the use of dimensionality reduction emotional features to the classification model training,can effectively improve the recognition rate of the system.Deep belief network as a deep learning structure,it is characterized first by the use of greedy unsupervised learning to train the network layer by layer,so as to get a good starting point,and then the use of supervised learning to optimize the whole deep structure.The network has strong capability of data abstraction and classification ability.Therefore,in the end of this paper,we use the deep belief network for speech emotion recognition experiment,and get a effective classification results on the same data set.