ИССЛЕДОВАНИЕ ЗВУКА: УДАЛЕНИЕ ШУМОВ

Abstract

The article presents algorithms for removing extraneous noise from an audio track. It considers the features of various types of noise that occur during sound recording. The article takes into account the features of the Conv-TasNet architecture, which is based on the imposition of convolutions on a pure signal without frequency separation. There is an analysis of the DEMUC algorithm, which directly generates sources from the original signal, bypassing the intermediate prediction of masks; the architecture for segmentation of images U-Net is partially borrowed. Also the authors consider the HiFi-GA noise reduction algorithm, consisting of three main parts: Wavenet, Postnet and GAN. A clean signal based on a noisy one is created with the WaveNet algorithm that was originally used to translate text information into speech. A feature of different versions of the WaveNet algorithm for noise reduction is that a new signal can be generated both in its entirety and for each time-point. The paper also presents a mathematical apparatus for implementing the ConvTasNet, DEMUC, and HiFi-GA algorithms, analyzes in detail noise reduction when recording sound, explores various noise reduction methods, and formulates the advantages and disadvantages of each of them.