Evaluating Lip-Reading Approaches: Comparative Study of Recent Techniques

Communication is necessary for human interaction. However, people with hearing or speech weakness face significant challenges in expressing themselves and understanding others. One technique to Get rid of these communication barriers is the Lip-Reading Technique, which involves translating silent videos into text. The lip-reading system includes various stages, starting from preprocessing, feature extraction, and finally classification. Lip reading started with traditional Machine Learning and improved with Deep Learning techniques. The expected output from the lip-reading system is typically in the form of sentences. These sentences may be constructed by combining various element shapes, such as ASCII Characters, Words, or Visemes. This paper provides a comparative study of two accurate lip-reading techniques and evaluates their performance. These approaches are “VSR for Multiple Languages in the Wild” and “AUTO-AVSR” which rely on Transformer, Conformer, and ResNet architectures. To perform this comparative study, two of the best widely used English sentence datasets for addressing lip-reading systems, LRS2 and LRS3, will be used. This comparative study also provides insights about which approach is suitable for building a bigger and more accurate lip-reading system for the Arabic language.