Abstract
The application of machine learning models for sign language recognition (SLR) is a well-researched topic. However, many existing SLR systems focus on widely used sign languages, e.g., American Sign Language, leaving other underrepresented sign languages such as Norwegian Sign Language (NSL) relatively underexplored. This work presents a preliminary system for recognizing NSL gestures, focusing on numbers 0 to 10. Mediapipe is used for feature extraction and Long Short-Term Memory (LSTM) networks for temporal modeling. This system achieves a testing accuracy of 95%, aligning with existing benchmarks and demonstrating its robustness to variations in signing styles, orientations, and speeds. While challenges such as data imbalance and misclassification of similar gestures (e.g., Signs 3 and 8) were observed, the results underscore the potential of our proposed approach. Future iterations of the system will prioritize expanding the dataset by including additional gestures and environmental variations as well as integrating additional modalities.