Running summarizing tools on a new article

Abstract: Sign Language Recognition (SLR) is critical in ensuring effective communication between hearing-impaired communities and hearing people. Current methods largely rely on time-consuming RGB-based systems, making their practical implementation and deployment limited in resource-constrained environments. In this study, we propose HBT-S2S, a new hybrid bi-directional LSTM (BiLSTM)-Transformer model for recognizing signs using only low-cost 2D pose, hand, and facial landmarks. What makes this method unique is the ability of capturing both local and global movements in a lightweight structure without the use of costly visual features. The stacked bidirectional LSTM captures small-scale dynamics, while the multi-head Transformer Encoder learns long-distance dependencies through self-attention. Also, we propose a sliced positional embedding approach that encodes motion phase relative information within a sequence and thus ensures robustness towards the variation in signing speeds - a feature largely ignored in previous studies. Furthermore, to ensure that padding tokens do not affect the accuracy of the output sequence, masked averaging pooling is employed to eliminate padded frames in the input sequence. HBT-S2S achieves Top-1 accuracies of 88.11%, 85.71%, 84.52%, and 83.42% on WLASL100, WLASL200, WLASL300, and WLASL1000. Additionally, an easy-to-train text-to-speech converter is incorporated to develop an end-to-end sign-to-speech pipeline.