From CNNs to Foundation Models: A Review of Modern Architectures for Visual Speech Recognition

Visual Speech Recognition (VSR), also known as Automated Lip Reading (ALR), poses a significant challenge in machine learning: the extraction of linguistic information from complex, high-dimensional spatiotemporal visual input. This paper offers a thorough architectural analysis, tracing the progression from traditional pipelines that were dependent on manually produced geometric or appearance-based features and statistical classifiers, to the modern framework of end-to-end deep learning. This review entails a thorough, component-based examination of contemporary VSR systems. We analyze the main visual front-ends, including ۲D/۳D Convolutional Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Vision Transformers (ViTs), assessing their techniques for collecting hierarchical spatial characteristics from the lip region. We present a comparative analysis of temporal back-end architectures, including Recurrent Neural Networks (RNNs), Temporal Convolutional Networks (TCNs), and Transformers, emphasizing their ability to model the intricate short- and long-range dependencies present in continuous speech. This paper provides a comprehensive path for researchers and practitioners by analyzing performance benchmarks, computational trade-offs, and fundamental concepts of these diverse architectural families. We conclude by synthesizing key architectural trends and proposing future research directions, focusing on efficient hybrid models and the deep integration of large language models.

کلیدواژه ها:

Visual Speech Recognition ، Automated Lip Reading ، Vision Transformers ، Large Language Models ، Spatiotemporal Modeling