Abstract:
Lipreading focuses on developing efficient models to decode spoken words from the motion of the lips. Despite continuous advancements, there remain two significant challenges: Model Complexity and Lip Movement Homophones. Complex models are often designed with the aim of increasing lipreading accuracy; however, this complexity comes at the cost of high computational resource requirements, rendering them unsuitable for use in portable devices or resource-constrained systems. Additionally, many words exhibit similar pronunciations, leading to lip movements that closely resemble one another; for instance, the words "million" and "millions" demonstrate such similarities, necessitating that models discern differences based solely on visual data. To address these issues, this research focuses on developing a streamlined yet highly accurate model capable of distinguishing between lip movements that closely resemble one another. This is achieved through the use of a hybrid deep learning approach that integrates Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Attention mechanisms, with the aim of enhancing the applicability and efficiency of lipreading in real-world scenarios.