Development and Research of Neural Network Based Method for Reconstructing Audio Signals

Abstract

In this paper we investigate the possibility of using neural networks to solve the problem of restoring audio signal. Based on the previously obtained results of the convolutional neural networks application for the extraction of a vocal part, we developed the concept of a convolutional neural network designed to correct distorted audio signal. The paper presents the initial concept of this neural network architecture which, unfortunately, showed unsatisfactory results. Nevertheless, based on the concept of this network, several new neural network architectures were developed specifically focused on recovering a distorted audio signal but the shortcomings of the basic architecture were taken into account. The paper contains descriptions of all these architectures and the results of their application to restore the drummer's part in the musical composition where it was removed. The obtained results show the high potential of convolutional neural networks application for solving such a complex problem as audio signal restoration.