Analysis of Lip Reading of Assamese Digits using Deep Learning

Effective communication in noisy environments, such as aviation, construction, and manufacturing, is often hindered due to auditory challenges, making oral communication difficult. To address this issue, we propose an automatic lip-reading system specifically designed for recognizing Assamese digits in high-noise settings. This study introduces a deep learning-based approach that extracts the geometric features of lip movements from video data to accurately predict spoken digits. Traditional lip-reading models struggle with language-specific nuances due to reliance on generic datasets. To overcome this limitation, we construct a custom dataset of video recordings featuring diverse speakers varying in age, gender, and accent, ensuring a more robust and adaptable model. We employ a CNN+LSTM architecture, where Convolutional Neural Networks (CNNs) capture spatial features, and Long Short-Term Memory (LSTM) networks learn temporal dependencies. Experimental results demonstrate that our CNN+LSTM model outperforms conventional architectures like RNN+LSTM and RNN+CNN, achieving an accuracy of 83%. The findings highlight the effectiveness of deep learning in enhancing accessibility for the deaf and hard-of-hearing and enabling voice-free human-computer interaction.