Year: 2025
Type: article
Abstract: The deaf and hearing-impaired community in Laos faces significant communication barriers due to the scarcity of resources and technological support for Lao Sign Language (LSL).This study addresses this gap by developing and rigorously evaluating a deep learning-based system for automated LSL recognition.We introduce a new, high-quality video dataset of 60 common LSL signs relevant to Lao culture and daily life.A comprehensive comparative analysis was conducted on four deep learning architectures: a CNN-LSTM, a CNN-LSTM with an attention mechanism (Hybrid-Attention), a Time Distributed CNN with Global Pooling (TD-CNN-Global Pool), and a 3DCNN-LSTM.All models were trained on a dataset expanded through extensive offline augmentation, using a 5-fold Group K Fold cross-validation strategy to ensure robust evaluation.The results demonstrate the critical importance of transfer learning, with models using a pre-trained MobileNetV2 backbone significantly outperforming a model trained from scratch.The standard CNN-LSTM architecture emerged as the superior model, achieving a mean cross-validation accuracy of 92.27%.The finalized model achieved an official accuracy of 93.33% on a completely unseen hold-out test set.This work establishes a strong baseline for LSL recognition, provides a valuable public dataset, and highlights the essential role of transfer learning and data augmentation in developing assistive technologies for low-resource languages.
Institution: National University of Laos
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: gold
OpenAlex ID: W4413813604