Study On Speech Emotion Recognition And Its Application

With the development and the increasing demand of Human-Computer Interaction (HCI) technique, affective computing has become a hotspot in the field of the current natural HCI and artificial intelligence. As one of the most direct and natural means of communication between people, speech is an important way of HCI. Not only semantic information, but also the speaker’s emotional state information can be conveyed by speech signal. Emotional communication is an indispensable part in human communication activities. Therefore, as an important direction of affective computing research, speech emotion information processing is drawing more and more attention of researchers. It is of great significance to research speech emotion recognition, which is the key to help computer understand human emotions.In order to establish a speech emotion recognition application system, in this thesis, we focus on emotional features analysis and extraction, emotional features selection, the speech emotion modeling and recognition. The main research work is as follows:(1) Front-end analysis and processing of speech signal. The methods of analysis and processing on time-domain waveform of speech signal have been introduced in detail. It mainly covers speech signal preprocessing, time-domain analysis and the principle and process of endpoint detection algorithm based on the combination of short-time energy and zero-crossing.(2) Emotional features analysis and extraction. An emotional speech database containing three different emotions of pleasure, calm and boredom has been established to solve the problem of data resource required in the speech emotion recognition research such as emotional features analysis and extraction. The research task of emotion features analysis and extraction has been carried out aimed at the emotions mentioned above. Variations of energy, pitch and formant of different emotions have been observed and analyzed. Otherwise, the global statistics emotional features with capacity of emotion identification have been extracted based on the result of statistical analysis.(3) Emotional features selection. In order to maximize the recognition speed of system and ensure that the system recognition rate is not significantly reduced, this thesis presents two schemes respectively based on emotion corpora (contains all three emotions) and emotion pairs(contains two emotions, three pairs). Through repeated experiments which apply the algorithms of Sequential Forward Selection and Sequential Backward Selection based on SVM error rate to the above two schemes, we get the optimal feature vectors from extracted emotional feature vectors. Then the comparison experiment of recognition performance of SVM emotion recognition model trained from the optimal feature vectors under different schemes on the same emotion test set has been carried out. The experimental results show that, SVM model trained from optimal feature vectors selected under scheme of emotion pairs has higher classification capacity and faster recognition speed on different emotion pairs.(4) Design and implementation of speech emotion recognition application system. In this thesis, a speech emotion recognition system applied under medical operation environment is designed and implemented. As a sub-module to be embedded in the developed minimally invasive surgery voice control robot system, the speech emotion recognition system is mainly used to solve two problems:First, the training phase, it helps doctors keep calm in training to improve performance; second, in the recognition phase, it will give decision aids when receiving the doctors’control commands to avoid the malfunction caused by doctors’mood swings and improve operation success rate.