Abstract: A mouth outline feature extraction based on Bayesian Tangent Shape Model (BTSM) and a lip-reading system based on Dynamic Bayesian Network is proposed for a talking head in this paper. This model describes the relationship of the word and its corresponding composed viseme, as a result, viseme segmentation sequence with time boundary is achieved. As a comparison, a DBN model based on word-phone relationship and a tri-phone HMM are used. For the system evaluation, an absolute Viseme Segmentation Accuracy (VSA) and two relative VSA based on image and geometrical feature of lip are brought out. The experiments show that DBN model has the better performance than HMM, and DBN model based on viseme can provide the best mouth shape for talking head.