BenAV: A Bengali audio-visual corpus for visual speech recognition

Visual speech recognition (VSR) is a challenging task. It has many applications such as facilitating speech recognition when the acoustic data is noisy or missing, assisting hearing impaired people, etc. Modern VSR systems require a large amount of data to achieve a good performance. Popular VSR datasets are mostly available for the English language and none in Bengali. In this paper, we present a large-scale Bengali audio-visual dataset, named "BenAV". To the best of our knowledge, BenAV is the first publicly available large-scale dataset in the Bengali language. BenAV contains a lexicon of 50 words from 128 speakers with a total number of 26,300 utterances. We have also applied three existing deep learning based VSR models to provide a baseline performance of our BenAV dataset. We run extensive experiments in two different configurations of the dataset to study the robustness of those models and achieved 98.70% and 82.5% accuracy, respectively. We believe that this research provides a basis to develop Bengali lip reading systems and opens the doors to conduct further research on this topic.