Feature Smoothing and Frame Reduction for Speaker Recognition

This paper presents a new technique for smoothing and reducing speech feature vectors for speaker recognition using an adaptive weighted-sum algorithm, aims at reducing computation time and increasing the recognition performance. The proposed technique is based on a three-frame sliding window. Each step of window sliding, three feature frames in the window are used to compute weight values based on feature Euclidean distances. The weight values are applied to original MFCC feature vectors to construct smoothed feature vectors. Simultaneously, the number of smoothed vectors is reduced from the original vectors. The smoothed and reduced feature vectors are applied on an SVM speaker recognition system with GMM super vectors. The NIST Speaker Recognition Evaluation 2006 core-test is used in evaluation. Experiment results show that our approach outperforms the baseline system using conventional RASTA filtered MFCC feature vectors.