An approach to vowel recognition using 2DDWT based visual information of the lip region

In this paper, a vowel recognition scheme using visual information is proposed based on two dimensional discrete wavelet transform (2D-DWT). First, a video frame corresponding to a steady vowel zone is selected utilizing the speech characteristics of audio frames. Next, a pixel-based method is proposed to identify the lip region of a given video frame, where intensity variation of different color planes is utilized. The 2D-DWT is then employed on a combined image plane extracted by using the weighted sum of red and green plane pixels of the lip image. Lower order wavelet coefficients obtained after second level decomposition and differences among those coefficients are used as proposed features. Leave one out cross validation technique is used to test the classification accuracy where a distance based classifier is used. Performance of the proposed method is tested on a publicly available standard audiovisual database and a high level of recognition accuracy is achieved using only extracted visual features.