Generative AI-assisted Standalone Wearable Reading Glasses for the Visually Impaired

This paper presents a novel, low-cost, wearable Optical Character Recognition (OCR) reading assistant glasses designed for the low-vision and visually impaired community. This standalone system uses a Radxa Zero 3W single board computer and a 5MP RGB camera. PaddleOCR engine, optimized through a series of preprocessing and postprocessing algorithms, used for effective text recognition. A finetuned lightweight 0.5B Large Language Model deployed on-device as a Generative AI chat assistant (Ask-LLM) to respond to the user query with full privacy. To convert the generated response into speech, the Festival Text-to-Speech (TTS) engine has been used. Also, Whisper CT2 Automatic Speech Recognition (ASR) has been utilized for effective and real-time voice transcription. The system recognizes English text with character recognition accuracy of 95.25% and Ask-LLM generates responses at 2~4 tokens/second. The hardware cost of the whole system is only $50 while works completely offline utilizing cutting-edge technologies. This paper contributes to the field of assistive technology by showcasing how carefully selected off-the-shelf components and open-source software can be engineered into a powerful, affordable, and practical solution.