Data Augmentation using GAN for Sound based COVID 19 Diagnosis

Abstract

The COVID 19 virus has been mutating at a rapid phase, due to which the golden standard of testing reverse transcription-polymerase chain reaction (RT-PCR) has been producing false negatives at an alarming rate. The inability of the test to detect the mutated strain of the COVID 19 virus using RT-PCR has made it very difficult for diagnosis and hence an alternative solution is needed. Sound-based diagnosis is one effective alternative diagnosis tool. The lack of a large dataset is one challenging aspect for the development of a sound-based diagnosis tool. We look forward to using dataset augmentation as a very effective technique for a selected classification problem: visual perception and also speech recognition tasks. The Generative Adversarial Networks (GANs) have been showing high success for applications in terms of synthesizing realistic images, they're seen rarely in audio generation-based applications Due to the lack of data sets available to develop an accurate model in this paper we showcase an application of WaveGAN, which is a variant of GAN which helps in raw audio synthesis during a supervised setting for the classification task, by developing a method showcasing one of the approaches for augmenting speech datasets by using Generative adversarial networks (GANs). We deploy the WaveGAN on the existing data sets collected from open-source collections to develop synthetic, larger data set to build an accurate sound-based diagnosis tool.