A Real-Time Sign Language Translator on Jetson Nano Using Mediapipe, MobileNet-V2 and LSTM
Md. Yusuf, Takia Nawrin, Maftun Ahommed, Nafees Mansoor
Conference Paper. IEEE International WIE Conference on Electrical and Computer Engineering/WIECON-ECE, no. 2025, pp. 1179–1182 (2025).
Abstract
The inability to communicate verbally isolates millions of deaf and hard-of-hearing people worldwide. The World Health Organization reports that more than 466 million people worldwide have disabling hearing loss. Sign languages are the primary means of communication for these people, but the lack of widespread sign-language literacy creates significant barriers in social, educational and professional settings. This article presents a complete design, implementation and evaluation of a real-time sign-language translation system running on an NVIDIA Jetson Nano. The system captures hand gestures with an on-board camera, extracts 3-D hand landmarks using Mediapipe Holistic, produces compact spatial features via a pre-trained MobileNet-V2, models temporal dynamics with a Long Short-Term Memory (LSTM) network and converts recognized signs into text. Comprehensive evaluation demonstrates that the proposed system achieves 98% accuracy on American Sign Language (ASL) letters, runs at 25-30 frames per second on Jetson Nano and consumes under 10W of power. The combination of efficient hardware and lightweight deep learning makes the proposed system a practical bridge between sign-language users and the broader population. © 2025 IEEE.
Keywords
Jetson Nano, LSTM, Mediapipe Holistic, MobileNet-V2, real-time translation, sign language recognition