Real-Time AI-Based Smart Hearing Protection Prototype for Enhancing Safety Signal Audibility in Work Zones

Abstract Hearing protection devices are essential for protecting the health of construction workers in job sites with loud noise; however, they may hinder workers’ ability to detect safety-critical alarms. Although previous research has demonstrated the potential of generative artificial intelligence (AI) models to enhance safety-critical sounds in construction work zones, the feasibility of real-time implementation of such models on wearable AI devices remains unclear. This study addresses that gap by developing and testing a novel AI-augmented hearing protector (AI-AHP) that deploys a real-time safety sound enhancement model on a low-cost Raspberry Pi 5 platform. The model was trained using various machine learning approaches on a data set comprising 4 types of safety-critical warning signals and 38 types of background noise. Controlled functionality testing, field deployment at active construction sites, and a subjective listener survey were conducted to evaluate the effectiveness of the prototype. The study identified DeepFilterNet as the most efficient model for edge computing, achieving fast and substantial improvements in signal audibility, with a maximum Perceptual Evaluation of Speech Quality (PESQ) improvement of 179%. Real-time operation was confirmed with a real-time factor of less than 1 and an end-to-end latency of 0.69 s on the Raspberry Pi 5. Subjective evaluation results showed significant increases in listener ratings, indicating that the enhanced signals were consistently perceived as clearer and more intelligible. These findings demonstrate the feasibility of deploying AI-driven hearing protection devices that improve worker situational awareness, reduce accident risk, and maintain usability in noisy construction environments, with potential applicability to other industrial settings.