Research On Feature Selection And Construction In Emotion Speech Recognition

Emotion Speech Recognition, which is a product of the interdisciplinary of artificial intelligence, pattern recognition, speech signal processing and affective computation, is a hotspot for researchers in recent years and has many importance application prospects in areas such as intelligent machines, human-computer interaction, remote teaching, criminal investigation and design, etc.Firstly, this thesis introduces the basis theory and methods of emotion speech recognition, and also introduces the most general emotion speech features and their extraction methods. Then, the statistic theory based classifier, namely the Support Vector Machine (SVM), is use for classification and output a good accuracy. After that, feature selection and feature construction are used to improve the accuracy.Feature selection is a very importance problem of pattern recognition. The thesis introduces the current status of feature selection and three basis algorithms, namely Sequential Floating Forward Selection (SFFS), ReliefF, Genetic Algorithm based Feature Selection(GAFS), is researched in detail. The Comparison result shows that GAFS, which can yield better features in fairy good period of time, is the best among the three. The features selected by GAFS improved the accuracy effectively.To further compress the feature space and improve accuracy, feature construction is researched. Feature construction is a novel pattern recognition technique, which maps present feature into a new function spaces thus can find relationships between features, and compress feature space, so as to improve the recognition accuracy. This thesis introduces existing algorithms, among which, Gene Expression Programming (GEP) based feature construction is better. Then a new feature construction algorithm that combines the Shuffled Frog Leaping Algorithms (SFLA) with GEP is proposed. The experiment, which uses the features selected by GAFS, that compares the SFLA-GEP and original GEP feature construction shows that SFLA-GEP beyond GEP in both convergence speed and convergence solution. The features constructed by SFLA-GEP can make effective improvement for emotion speech recognition accuracy.