Year: 2025
Type: book-chapter
Abstract: Lesotho Sign Language (LSL) remains an underrepresented language with less than 4500 speakers, unavailable datasets and limited assistive technologies.This study presents the development of an LSL dataset comprising gesture-based images of alphabet letters A-Y, collected and annotated in collaboration with the National Association of the Deaf Lesotho.A Convolutional Neural Network (CNN) based model was developed for the Sign Language Translation (SLT) task.Then, a hierarchical data format version 5 (HDF5) version of the trained model was used to translate hand gestures to text and audio for the Sign Language Recognition (SLR) task.The model achieved 92.61% accuracy in the SLR task, and promising results in the SLT task for letters N, C, H, O and D. This work provides a foundational framework for empowering the speech and hearing-impaired community in Lesotho in communicating and accessing crucial services like healthcare and education through assistive technologies.Future research could explore the integration of motion-aware architectures and Contrast Limited Adaptive Histogram Equalisation (CLAHE) preprocessing to enhance model performance under varying lighting conditions and to better handle sign variations across the community.
Authors: William Nkomo, Refuoe Mokhosi, John Batani
Institution: Botho University
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4417192533