Adaptive and Optimal Classification of Speech Emotion Recognition

It is important to properly select and extract the features of speech emotion, and to reasonably construct the classifier for improving the accuracy of the speech emotion recognition. In this paper, the cubic spline fitting is used to fit curves of prosodic features extracted from speech signals and then the derivative parameters features of these fitting curves are attained. We closely combined the stage of feature selecting and the stage of feature classification, and considered the personal characters of different emotions based on genetic algorithm (GA) and support vector machine (SVM) classification algorithm. Using the optimal searching property of the GA, the system attained the maximum recognition rate by adaptively searching the order of emotion selection and the selection subset of features. This system's average recognition rate can reach as satisfying as 88.15% over six emotions.