Abstract: Automated recognition of sign languages has been playing an important role in breaking barriers to communication and inclusion for the deaf and mute community. Several studies have been conducted on Bengali Sign Language (BdSL). However, Bengali Word-Level Sign Language (BdWLSL) remains unexplored due to the lack of large annotated datasets and a stable model. Therefore, in this research, we introduced a large-scale Bengali word-level video dataset and proposed a modified 3D Convolutional Neural Network (CNN) architecture for word-level BdSL recognition, emphasizing its ability to capture the spatial and temporal dynamics from video data. The proposed strategy represents strong performance in Bengali word-level sign language recognition by utilizing the spatiotemporal pattern captured by the modified 3D CNN architecture. The proposed model demonstrates its potential for practical use by successfully learning complex hand movements straight from raw video data. The proposed CNN model is benchmarked against traditional deep learning techniques, Temporal Shift Module (TSM), Long Short-Term Memory (LSTM), and default 3D-CNN, providing a comprehensive comparison of their strengths and limitations. Experiments are conducted using a structured video dataset containing 102 Bengali sign-word classes. To ensure privacy, the volunteers’ faces were blurred and only landmark data extracted using MediaPipe, rendered on black backgrounds, were used for training. The experimental result analysis shows that the performance of the proposed 3D-CNN model achieves a satisfactory accuracy of 58.25%, demonstrating its potential for word-level sign language recognition tasks. To our knowledge, this is the very first pilot study for BdWLSL recognition. Hence, we consider the recognition rate 58.25% of the proposed modified 3D-CNN architecture to be satisfactory and a potential scope for future researchers in the same field.

Safi Ullah Chowdhury, Nasima Begum, Tanjina Helaly and Rashik Rahman. “Sign3DNet: An Enhanced 3D CNN Architecture for Bengali Word-Level Sign Language Recognition”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.6 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0160689

@article{Chowdhury2025,
title = {Sign3DNet: An Enhanced 3D CNN Architecture for Bengali Word-Level Sign Language Recognition},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2025.0160689},
url = {http://dx.doi.org/10.14569/IJACSA.2025.0160689},
year = {2025},
publisher = {The Science and Information Organization},
volume = {16},
number = {6},
author = {Safi Ullah Chowdhury and Nasima Begum and Tanjina Helaly and Rashik Rahman}
}