Intensity/location normalization for automatic lipreading

This paper describes intensity and location normalization for the improvement of the performance of a speech recognition system by using the visual information in bimodal speech recognition. In conventional speech recognition, many methods have been proposed for normalization of channel characteristics and speaker individuality. In this study, two methods similar to CMN and SAT were proposed for the intensity and location normalization respectively, and two kinds of feature vectors, a subsampled image and a 2D-DCT were compared. Experimental results show that the recognition rates have been very much improved by the normalization techniques