A wearable glove-based system for real-time sign language recognition using parallel hybrid CNN–LSTM

Abstract

Real time sign language recognition (SLR) is essential for bridging communication gaps for the hearing impaired; however, existing solutions often suffer from high hardware complexity, excessive weight, and processing latency. To address these limitations, this paper presents a lightweight, cost-effective wearable gesture recognition system, evaluated on a 12-class constrained gesture vocabulary as a step toward scalable continuous SLR, integrating a custom sensorized glove with a parallel hybrid convolutional neural network (CNN)–long short-term memory (LSTM) architecture. Unlike conventional tandem pipelines, the proposed model extracts multi-scale spatial features via CNNs and inter-channel (cross-sensor) dependencies via LSTM units concurrently, thereby significantly enhancing computational efficiency. The wearable hardware, weighing only 64 g, utilizes five flexible strain sensors and a single MPU6050 inertial measurement unit to capture 11-dimensional hand-motion data. To ensure mechanical reliability, finite element analysis of the sensor substrate was performed under cyclic bending conditions. Experimental results on a 12-class gesture dataset demonstrate that the system achieves 98.69% accuracy on controlled test sets, independently confirmed at 98.82% under a matched-ratio chronological train/test split verifying robustness against evaluation-protocol bias, 95.49% accuracy in user-dependent real-time tests, and 92.24%–95.21% accuracies in user-independent real-time evaluation using continuous-stream gesture monitoring. Consequently, the proposed framework outperforms several sensor dense state-of-the-art systems, proving that high-fidelity gesture translation can be achieved with minimal hardware and low latency parallel processing. These results establish a practical and scalable solution for assistive communication and human–computer interaction.