Towards a Lightweight Arabic Sign Language Translation System

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Effective communication between deaf and non-deaf individuals is crucial.
• Existing sign-to-text systems often lack the performance or efficiency for real-time applications.
• Saudi Sign Language (SSL) datasets are essential for developing specialized translation tools.

Purpose of the Study:

• To develop a high-performance, lightweight sign-to-text translation system for real-time applications.
• To evaluate the system's effectiveness using two Saudi Sign Language datasets (KSU-SSL and ArSL).
• To investigate the impact of dataset size (number of signers and repetitions) on translation accuracy.

Main Methods:

• Implementation of a sign language translation model evaluated in signer-dependent and signer-independent modes.
• Conducting eight experiments to assess various configurations and dataset parameters.
• Performing a comprehensive ablation study on model components, network depth, and hidden dimension size.

Main Results:

• The system achieved 97.7% accuracy in signer-dependent mode and 90.7% in signer-independent mode on the KSU-SSL dataset.
• On the ArSL dataset, the model reached 98.38% accuracy (signer-dependent) and 96.22% (signer-independent).
• The study analyzed the influence of dataset characteristics and model architecture on performance.

Conclusions:

• The developed sign-to-text system demonstrates high performance and efficiency for real-time Saudi Sign Language translation.
• The findings highlight the importance of dataset size and signer variability in model training.
• This research contributes to improving communication accessibility for the deaf community through advanced AI.