LipWiz: A Proposed Approach towards Visual Speech Recognition

The act of lip reading, or visual speech recognition has attracted more researchers in the last few years as it can be employed in human computer interaction, enhancing accessibility and communication without sounds. This paper reviews different mechanisms that are designed for lip reading with respect to analysis of Lip-based Word Identification Zone (LipWiz) in the first place. In comparison with LipNet, some modifications are made in LipWiz, such as providing new datasets and replacing Bi-directional Gated Recurrent Units (GRUs) with Bi-directional Long Short-Term Memory networks for enhancing temporal sequence modelling. Moreover, LipWiz has been benchmarked with other recent lip reading architectures to show its improvements in accuracy, generalisation, and even computational performance. LipWiz exhibits far better accuracy, generalization, and performance in computation as compared to other architectures such as LipNet.