Cough Sound Analysis for the Evidence of Covid-19

Nicholas Rasmussen, Daniel L. Elliott, Muntasir Mamun, KC Santosh

Abstract: Efficient techniques for Covid-19 screening tests are crucial to preventing infection rates. Although symptoms present differently in different socio-demographic groups, cough is still ubiquitously presented as one of the primary symptoms in severe and non-severe infections alike. Audio/speech processing is no exception in health care. In this paper, we implemented a convolutional neural network (CNN) algorithm to analyze 121 clinically verified cough audio files from the volunteer group ‘Virufy’ for Covid-19 screening. Using a single relatively small CNN with a large, fully connected dense layer trained on melspectrograms alone, we achieved 0.933 test accuracy and an AUC of 0.967 on a small dataset. Our results are competitive with state-of-the-art results with both small and large datasets.