Wind Noise Reduction Strategy in Hearing Aids Through U-Net Deep Learning and Microphone Enclosure Design

Wind noise interference can adversely affect the performance of behind-the-ear hearing aids, particularly when used outdoors. In this study, we developed a novel strategy to mitigate this problem. Our methodology involved a modified microphone enclosure design, preprocessing stages of beamforming and wide dynamic range compression (WDRC) (which are inherent in conventional hearing aids), and advanced deep-learning-based noise reduction methods. We first explored the cavity aspect ratios and wall slanting of the microphone enclosure's design and determined that noise could be reduced by modifying the upper and lower enclosure widths from 0.8 mm (for both) to 1.1 and 0.5 mm, respectively. In terms of signal processing, a 1-D convolutional neural networks (CNNs) model achieved wind noise detection with an accuracy of 99.25% in various scenarios. The U-Net deep learning architecture was implemented for noise reduction and substantially improved short-time objective intelligibility (STOI) by 18.97%-209.09% compared with traditional high-pass filters (HPFs). Training with a voice database further improved the STOI. In terms of the mean opinion score-listening quality objective (MOS-LQO) and STOI metrics, all combinations of preprocessors with U-Net outperformed U-Net alone, and beamforming was the optimal preprocessing method. In conclusion, adaptive signal preprocessing based on wind classification, microphone enclosure optimization, and U-Net deep learning techniques effectively reduced wind noise interference, improving the outdoor usability, and listening experience provided by hearing aids.