Noise Cancellation and Speech Enhancement Using Multi-Layer Convolutional Neural Network

Abstract

We are introducing a new end-to-end deep learning model which aims to cancel noise in the background and enhance the speech simultaneously. The input to the model is given as a synthetically produced Noised-Speech signal by mixing clean speech signals and noise signals. The aim is to get a denoised speech with the noise removed from the background. In our approach we used Wave-U-Net as the base model to draw an idea upon the final model. The framework used is PyTorch along with Fastai for training and inferencing. The loss function used in the paper is a modified MSE loss.