Visual Speech Recognition in Bangla: A Convolutional LSTM Framework

Lip-reading technology enables speech interpretation through visual analysis of lip movements, serving as a crucial tool for communication in noisy environments and for hearing-impaired individuals. Although Bangla is spoken by approximately 265 million people worldwide, automated lip-reading systems for this language remain underexplored, primarily due to the lack of comprehensive datasets and specialized models. To address this research gap, we present a novel Bangla lip-reading dataset comprising video recordings of native speakers articulating common Bangla words. Our dataset ensures diversity through varied speaker demographics and controlled recording conditions. We feature a convolutional-LSTM architecture leveraging robust feature extraction and time-series feature analysis. Our model surpasses other state-of-the-art data-driven methods in terms of various performance metrics. Ablative analysis in LIP-NET proves the superiority of our solution over various experimental settings.