Duration Normalization Algorithm Based on Feature Space Trajectory in Pathological Speech Recognition

In this paper, a Duration Normalization algorithm is presented based on feature space trajectory to improve the accuracy in pathological speech recognition. In this method, speech MFCC features are divided into K segments dynamically, then the intra-segment mean values of each K segment are used to be new features. Our approach makes the difference (dispersity) between different segments the largest, and the difference between the feature vectors within the same segment is the smallest. Therefore, the redundant information and the distortion phenomenon of the speech signal could be removed. Experiment shows that in speech recognition system, using our approach, the accuracy increases by 11%-15% and the run time decreases by 14%-40%.