Speech Enhancement using Deep Convolutional Neural Network

Abstract

This work presents a supervised speech enhancement method using a deep convolutional neural network (CNN). The proposed CNN is based on a Convolutional Autoencoder architecture with symmetric skip-connections. Additionally, we focus on building a novel and robust dataset for this task. The data contains a clean speech dataset and a noise dataset, and each outweighs its counterpart used in recent works. Finally, we investigate the performance of the system on many levels of noise by performing the evaluation using objective metrics that are commonly used in this area.