Abstract
This paper presents a low-cost audio denoising system designed as an over-the-counter (OTC) hearing aid that does not require a prescription. The system uses a modified U-Net convolutional neural network with a ResNet-101 encoder, trained on Mozilla Common Voice speech data and ESC-50 environmental noise. Training data were filtered using the Hearing Aid Speech Perception Index (HASPI) to retain only file pairs with a score of 0.6 or higher. The trained model was compressed using TensorFlow Lite and deployed on a Raspberry Pi 5, where it denoises audio received through a wireless headset. Evaluation on 51 unseen files showed consistent improvement in both Signal-to-Noise Ratio (SNR) and HASPI scores, with most denoised files achieving SNR values above 3 dB and HASPI scores above 0.9. At a total cost of approximately 50, the system demonstrates that an effective AI-based hearing aid prototype can be built with relatively inexpensive tools.