Generative Adversarial Networks for Respiratory Sound Augmentation

Abstract

In this paper we propose to use generative adversarial network (GAN) for respiratory sound data augmentation. We present a GAN based approach that requires moderate amount of time and computing resources and capable to greatly increase performance of lung sound classification tasks. We also present a conditioned version of GAN, which is flexible and outperforms competitor augmentation methods. As a result, the GAN based augmentation method is able to boost RNN classifier performance by 10-15