Kernel-based Speaker Verification Using Spatiotemporal Lip Information

The lip-region can be interpreted as either a genetic or behavioural biometric trait. Despite this breadth of biometric content, lip-based biometric systems are scarcely developed in the literature. A recent trend in lip biometrics is to use a spatiotemporal texture representation of visual speech to generate biometric features. In this paper we make two contributions related to the above biometric traits. We investigate whether the application of non-linear discriminant analysis on spatiotemporal texture improves its biometric performance. Spatiotemporal texture representation of visual speech results in performance that suggests that the lip can be used as a hard biometric. We investigate the effect of the amount of video information on speaker verification performance. The results show that using non-linear discriminant analysis improves speaker verification performance. Additionally, we also demonstrate that using over 3 seconds of video is sufficient to achieve satisfactory accuracy.