A Lightweight Solution for Pose-Based Recognition for Isolated Spanish Sign Language Using Recurrent Models

Abstract

This work addresses the problem of Isolated Sign Language Recognition (ISLR) in Spanish Sign Language (LSE) from a pose-based perspective. The proposed approach relies on 3D landmark extraction using Google’s MediaPipe framework to obtain face, hand, and upper-body keypoints, which are then normalized and transformed into spatial–temporal feature sequences. Two temporal alignment strategies, average sampling and max-length padding, were implemented to ensure uniform input dimensions across samples. Bidirectional recurrent neural networks (Bi-LSTM and Bi-GRU) were evaluated to capture the temporal dependencies inherent to signing. Experimental results on the LSE-Health-UVigo dataset show that the Bi-LSTM architecture combined with the Focal Loss function (γ = 3) achieved the highest performance, reaching 79.8% unweighted accuracy. The proposed model has an average response time of approximately 1 ms, making it suitable for deployment in real-time scenarios. These results highlight the effectiveness of pose-based recurrent architectures for ISLR and demonstrate the potential of lightweight models for robust sign language understanding.