U-Net-Based Single-Channel Wind Noise Reduction in Outdoor Environments

Geon Woo Lee, Kwang Myung Jeon, Hong Kook Kim

Abstract: This paper proposes a noise reduction method based on a U-shaped neural network to effectively reduce wind noise. While the U-Net is developed for medical image segmentation, it is constructed by using the spectrograms of noisy input signals as the input feature, and it is trained to estimate the ideal ratio mask between a pair of input noisy and clean target signals. The performance of the proposed method is measured in terms of signal-to-distortion ratio (SDR), signal-to-interference ratio (SIR), and signal-to-artifact ratio (SAR). As a result, it is shown that the proposed method provides a higher average SDR, SIR, and SAR than conventional statistical methods such as minimum statistics-based and nonnegative matrix factorization-based methods under various signal-to-noise ratio conditions.