LipReadNet: A Deep Learning Approach to Lip Reading
Kuldeep Vayadande, Tejas Adsare, Neeraj Agrawal, Tejas Dharmik, Aishwarya Patil, Sakshi Zod
Abstract
LipReadNet is a deep learning approach to lip reading that aims to improve speech recognition technology for individuals with hearing impairments or in noisy environments. Lip reading has long been known as an effective method of communication for people who have hearing problems, and with the advancement of deep learning algorithms, it has become possible to automate the process of lip reading. The LipReadNet model has the potential to revolutionize the field of speech recognition technology, making it more accessible and useful for individuals with hearing impairments, as well as in scenarios where the audio signal is degraded or absent. The LipReadNet model comprises a 3D CNN and LSTM that is trained on large datasets of video and audio recordings. The model first extracts visual features from the mouth region of a person’s face, then combines these features with the corresponding audio signal to predict the spoken words. This approach is highly effective as it can recognize spoken words even in cases where the audio signal is corrupted or missing entirely. LipReadNet outperforms existing lip-reading models in terms of accuracy, robustness, and efficiency. The goal of a lip-reading project is to develop a system that can accurately recognize speech through visual cues, without the use of audio. Achieved accuracy of 93% in a lip-reading project which rely on different factors, such as quality of data and diversity of data of training, the choice of machine learning algorithms, and the evaluation metrics used to calculate the competence of the system.