Resumen
This paper studies various deep learning models for word-level lip-reading technology, one of the tasks in the supervised learning of video classification. Several public datasets have been published in the lip-reading research field. However, few studies have investigated lip-reading techniques using multiple datasets. This paper evaluates deep learning models using four publicly available datasets, namely Lip Reading in the Wild (LRW), OuluVS, CUAVE, and Speech Scene by Smart Device (SSSD), which are representative datasets in this field. LRW is one of the large-scale public datasets and targets 500 English words released in 2016. Initially, the recognition accuracy of LRW was 66.1%, but many research groups have been working on it. The current the state of the art (SOTA) has achieved 94.1% by 3D-Conv + ResNet18 + {DC-TCN, MS-TCN, BGRU} + knowledge distillation + word boundary. Regarding the SOTA model, in this paper, we combine existing models such as ResNet, WideResNet, WideResNet, EfficientNet, MS-TCN, Transformer, ViT, and ViViT, and investigate the effective models for word lip-reading tasks using six deep learning models with modified feature extractors and classifiers. Through recognition experiments, we show that similar model structures of 3D-Conv + ResNet18 for feature extraction and MS-TCN model for inference are valid for four datasets with different scales.