On the use of VGGish as feature extractor for COVID-19 cough classification

The COVID-19 pandemic has changed the daily lives of all people worldwide, affecting not only society but also various sectors such as finance, tourism, etc. To counteract the pandemic, measures are required to detect contagions and take the necessary actions to prevent the virus spread. In this work, a Transfer Learning approach has been used to model COVID-19 coughs as a previous step to the diagnosis of the illness. The data set of the University of Cambridge, ComParE 2021 COVID-19 Cough Sub-Challenge, has been used, which consists of 725 samples of cough sounds from 397 people of which 119 have been diagnosed with positive COVID-19, besides, a data augmentation technique has been used to balance the data set. This work evaluates the performance of the pre-trained VGGish model for the classification of the audio cough signals as COVID or Not COVID cough. For this purpose, the VGGish model is used as a feature extractor and a convolutional neural network provides the final classification of the cough recordings to determine whether they are COVID-19 positive or negative. Despite the difficulty of the task, optimum results have been founded to detect negative cases obtaining up to 81% of precision. Considering the Unweighted Average Recall (UAR) as metric, the methodology proposed in this work has obtained an improvement up to 3% comparing to OpenSmile technique when the same database has been used.