There is number of researches on the lip reading. However, there is little discussion about which face model is effect for lip reading. This paper builds various face models which changes the combination of a face part, and changes the feature points. Various experiments were conducted on the conditions which change only model and do not change other algorithms. We apply the active appearance model. The CUAVE database which the utterance scene of ten digits is contained was used for the recognition experiment. As a result, the model which combined the external lip contour, nose, and outline acquired the highest recognition accuracy using HMM. We found that the face model which contains eyes and/or eyebrows is not effective for lip reading.

Index Terms: lip reading, word recognition, face models, active appearance model