DEEP LEARNING-BASED DYNAMIC SIGN LANGUAGE RECOGNITION MODEL

ABSTRACT Communication barriers between hearing- and speech-impaired individuals and the wider community continue to pose challenges to inclusivity. Sign language, while effective within its user community, often becomes a limitation in interactions with non-signers. This project focuses on the development of a deep learning–based dynamic sign language recognition model capable of translating sign gestures into textual output, thereby enhancing accessibility. The proposed system employs a hybrid architecture that integrates Convolutional Neural Networks (CNNs) for spatial feature extraction with Bidirectional Long Short-Term Memory (BiLSTM) networks for temporal modeling of video sequences. A dataset comprising 3,522 video samples of five American Sign Language (ASL) gestures—again, bad, bathroom, book, and busy—was collected from Kaggle and preprocessed through frame sampling, resizing, and normalization. The CNN–BiLSTM model was implemented in Python using TensorFlow, OpenCV, and Google Colab GPU resources. Experimental results demonstrated exceptional performance, with the model achieving 100% accuracy, precision, recall, and F1-score on both validation and test sets. These outcomes validate the effectiveness of the hybrid CNN–RNN framework for capturing spatiotemporal features in dynamic gesture recognition. While the findings establish a strong baseline, the study acknowledges key limitations such as the small dataset size, restricted gesture vocabulary, and evaluation under controlled indoor conditions only. The project contributes to the field of assistive technology by providing a reproducible framework for sign language recognition and highlights avenues for future research, including expanding the dataset, improving real-time deployment, and supporting multiple sign languages. Overall, this work underscores the potential of deep learning to promote inclusivity and bridge communication gaps for the hearing- and speech-impaired community.