Research On The Related Algorithms Of Speech Emotion Recognition

Speech emotion recognition is one of the research hotspots in affective computing, referring to the process of the preprocessing of the collected speech data, the extraction of the features that effectively characterize different emotions, the use of the classification algorithm to identify the underlying emotion in the speech. As the key of intelligent human machine interaction, it is widely applied in intelligent machines, call center system service, online education, medicine diagnosis and other fields. Numerous researchers have done researches on the speech emotion recognition. Remarkable progress has been achieved. But there are certain problems remain to be solved in how to acquire effective and discriminative emotion features and seeking for efficient and stable recognition algorithms.In this paper, related topic as the background, on the basis of speech data preprocessing and the extraction of emotion related feature, feature dimension reduction and speech emotion recognition are studied. The main works are presented as follows:The projection separating class from others obtained with the classical linear discriminant analysis method is not the best one due to outlier class. To approach the problem, a novel distance weighted linear discriminant analysis method for dimension reduction of emotion feature is proposed, which improves the separability of emotion speech in the new projection space.The majority voting mechanism of Random Forests fail to consider the differences of classification performance between different base classifiers, resulting in poor classification performance. To tackle this issue, a speech emotion recognition method based on improved voting mechanism is introduced. The base classifier predicts the class labels of OOB samples to get the classification accuracy rate which is treated as the voting weight. The method improves the voting weights of the powerful classifiers in order to improve the whole classification effect. The experimental results demonstrate the effectiveness of this method.When an attribute is independent of the class, using Gini index to choose the best attribute splitting the node leads to poor performance in Random Forests. In view of this, a speech emotion recognition method based on ReliefF and Gini index is presented. It increases the accuracy of speech emotion recognition.