Abstract: For the visual feature extraction of visual only lipreading system, this paper proposed an DCT}ONPP method to extract the visual feature of lip, compared with PCA which preserving the global structure, ONPP is a method preserving the intrinsic geometry of the local neighborhoods, in the same time, the overlap of the neighborhoods can keep the global structure of the data. The experimental results show that the proposed method is better performance than PCA and DCT+PCA methods. About the influence of the neighbor numbers on the system, some research has been done on it, the experimental results show that 3 neighbors can get better performance.

Key words: Lip feature extraction, ONPP, Lip reading, PCA