Speech Emotion Recognition Based On Deep Belief Networks

With the research on affective computing going in recent years, speech emotion recognition has attracted increasing attention from the researchers. It’s very helpful for the development of psychology and building more harmonious environment in the human-computer interaction. Speech emotion recognition extracts a vector that contains some emotion features, processes the vector through the classification model, and returns the result in the end. How to improve the recognition performance of classification model is always the key point.To improve the accuracy, this thesis proposes a strategy to recognize speech emotion based on the deep belief networks, which improves the generalization of complicated classification, reduces the training time of neutral network, and makes up the traditional neutral network’ s shortcomings on the selecting features and expressing complicated function. This thesis realizes the strategy via MATLAB, and carries out the experiment to compare the strategy to BP neutral network, in terms of emotion recognition of recall, precision and F1. The results show that the proposed strategy achieves higher recognition performance than BP neutral network in the metrics of the mean recognition recall, precision and F1of six emotions.Based on the above-mentioned strategy, this thesis proposes the prototype of a system and realizes it. The system is developed for Android smartphones, and adopts the C/S structure. The clients’ major function is to recording, while the server is for feature extraction and emotion recognition. Users record their speech through the microphone, and upload the files to the server. The server will analyze the speech and return the result of emotion classification.