Speech Enhancement using Convolution Neural Network-based Spectrogram Denoising
Xuhong Hu, Yan Lin-Huang, Xun Lu, Guan Yuan-Sheng, Wenlin Hu, Jie Wang · 2021
Regarding spectrogram as an image, this paper adopts a convolution neural network (CNN)-based image enhancement algorithm for spectrogram denoising. By doing so, speech denoising can be achieved when the spectrogram is enhanced by the proposed CNN-based image enhancement algorithm. The spectrogram clipping strategy was presented to obtain a large amount of training data, which gave rise to a smaller storage cost and avoided the limited depth development and problem of excessive complexity commonly presented in traditional speech features when training a recurrent neural network. Meanwhile, a deeper network was constructed to improve the capacity and flexibility to use the features of the spectrogram better, and it can also capture enough spatial information to make the noise reduction performance effectively. In addition, the proposed model utilized residual learning strategy in CNN training, with the combination of batch normalization, which greatly improved the performance of the model. The experimental results demonstrates that the proposed spectrogram denoising model has better learning ability and denoising performance, whether it is a known noise situation or a noise mismatch situation, so that the proposed system shows robust speech enhancement effect.