Uni-Modal Pipeline for Data-Agnostic Sign Language Recognition: Framework & Evaluation

The automatic recognition of continuous sign language presents a significant challenge in artificial intelligence due to its complex linguistic structures and the scarcity of standardized research tools. This paper introduces and evaluates a new uni-modal architectural pipeline for data-agnostic sign language recognition, prioritizing the analysis of manual gestures to balance linguistic completeness, computational efficiency, and user privacy. The system's core is a Bidirectional Long Short-Term Memory (Bi-LSTM) network, integrated into a scalable, low-latency inference system that uses client-side feature extraction and gRPC for efficient data transfer. Validated on a custom dataset with a constrained vocabulary, the hands-only pipeline achieved an impressive F1-Score of 0.978, confirming its effectiveness as a highly efficient baseline for future sign language recognition research. This performance suggests that for many real-world applications, a uni-modal approach can provide an optimal balance between accuracy and practical applicability, especially when user privacy and computational resource optimization are crucial. The framework's modularity also allows researchers to isolate and investigate specific aspects of sign language recognition, facilitating more targeted and efficient model development and dataset curation. This work highlights the potential of a hands-only approach to create accessible and privacy-preserving sign language recognition technologies.