Abstarct: Biometrics is a type of security measure for access control in computer systems that uses the measurable characteristics of the human body to authenticate the user, and these characteristics fall into two general categories, physiological and behavioral. In the last decade, the field of biometrics has undergone significant transformation and advancements thanks to deep learning. In this regard, many improvements have been applied to the old biometric methods, which have significantly reduced their security problems. One of the most widely used methods of image biometrics is using the image of people's faces, which despite its high safety Unfortunately, according to some people, it is considered a serious threat to privacy. Some other biometric methods that use audio features are also not always usable for reasons such as noise disturbances in different environments and lack of usability by people with speech disorders. Lip-baxsed biometry (LBBA) has attracted many researchers in the last decade because of the above problems. The lip is of particular interest to biometric researchers because it is a binary feature with the potential to function as a physiological and behavioral feature simultaneously. Although a lot of valuable research has been done on LBBA, most of them have not taken into account the different emotions of people during the LBBA imaging stage.which can potentially affect facial exxpressions and speech speed. In this thesis, a new neural network architecture is proposed that uses a Siamese deep structure with triple error function with three identical slow-fast networks as feature extraction networks. The slow-fast network is an excellent choice for performing feature extraction from lip images because the fast path of this network captures motion-related features (lip behavioral features) with a high frxame rate and low color channel capacity, and the slow path captures visual features (lip physiological features). ) with low frxame rate and high color channel capacity. The proposed network was trained using the CREMA-D databaxse and achieved an equivalent error rate (EER) of 0.005 on the test set.