Lip Reading Experiments for Multiple Databases using Conventional Method

Abstract

Several databases have been published in the field of lipreading. Each database has a wide variety of utterance content, number of speakers, shooting angles, and so on. However, each paper uses one or two databases. It is desirable to discuss using many databases. This paper uses four publicly available databases: CUAVE, OuluVS, CENSREC-1-AV, and SSSD. In this paper, not the latest deep learning-based method but the standard recognition method by hidden Markov model (HMM) which mainly used conventionally is applied, and analyzes trends in recognition accuracy. Based-on recognition experiments, it was found that the recognition accuracy was correlated with the number of frames.