Research On Key Technologies Of Human Identification Based Lip Reading

With the rapid development of science and technology, human's consciousness of security is improved gradually. The traditional security technology doesn't meet the requirements of security quality at present. Then, the people turn to multichannel recognition technology and make it develop fast.Now,the technologies of human identification based lip reading match result more effective for AVSR(Audio Visual Speech Recognition).In this paper, some key issues of lip reading technology have been studied .The main works as follows:Design a experimental platform for lip reading.The AVSR datdbase is Tulips1 and the nosie teset based on Aurora2.0. Besides,the platform has been implemented by OpenCV1.0 for image feature extraction effective.Design a test methods for experiment.In this paper,we analysed the methods of experiment very closely, especially the error estimation and the adding method of noise.Finally,use rotation error estimation.Extraction speech feature for lip reading.In this paper,we have compared three methods to extract speech feature in this study, including: MFCC(Mel Frequency Cepstral Coefficients);PLPC (Perception of Linear Prediction Coefficients);energy combine ZRC (Zero Crossing Rate).The conclusion is that MFCC is the best method on the database that we chosed.Design a multichannel recognition system for human identification.In this paper,a new DTW(Dynamic Time Warping) for recognition has been proposed.The result of experiment show that the new DTW is more effective.Both image feature and speech feature have been fused in this methods.