Speech denoising based on U-shaped neural network

Abstract

Due to the increasing need for speech enhancement applied to speech communications or automatic speech recognition, several denoising methods have been proposed, including sparsity nonnegative matrix factorization (SNMF) [1] and speech enhancement generative adversarial network (SEGAN) [2]. Recently, a U-shaped neural network (U-Net) composed of a series of convolutional neuralnetworks (CNNs) has been popular in various areas of signal processing, especially image processing [3]. This paper proposes a U-Net-based speech denoising method for enhancing noisy speech under low signal-to-noise ratio (SNR) conditions.