earticle

영어

Feature dimension reduction is important for speech emotion recognition. The classical linear discriminant analysis has been used widely in this field, but the best projection separating class from others can’t be obtained with the linear discriminant analysis method due to outlier class. To approach this problem, a novel distance weighted function based on the linear discriminant analysis is introduced, which can improve the separability of sample data and has low computational complexity. In order to evaluate the proposed algorithm’s performance, some experiments are performed on two speech databases: UCI and CASIA. Experimental results on the UCI database demonstrate that the presented algorithm has a better performance. Experimental results on CASIA show that the proposed algorithm yields an average accuracy of 88.78% in the classification of four emotions, revealing that it is a better choice as feature dimension reduction for emotion classification than the traditional algorithms.