Enabling COVID-19 Detection from Multiple Audio Recordings: A Preliminary Comparison Between Cough, Breath, and Speech Signals
摘要
Patients with COVID-19 experience severe respiratory and vocal difficulties as well as symptoms that give rise to unique audio characteristics in their voices. The present study takes advantage of vocal biomarkers extracted from cough, speech and breathing audio recordings obtained through personal smartphones from both SARS-CoV-2 virus-infected persons and non-infected participants accessing two different healthcare facilities. The results provide findings on the use of acoustic feature sets taken from low-level feature representations for COVID-19 recognition from cough, breath, and speech patterns. Machine learning models were trained on datasets from individual vocal exercises and on a dataset from combined exercises (cough, breath, and speech). The classification models provided up to 68.6% accuracy, 86.7% sensitivity, and 66% specificity, whose values and most significant features vary according to the type of vocal pattern examined and the type of model adopted, indicating that audio characteristics may be used to detect COVID-19 symptoms and that the combined use of multiple audio patterns from different vocal tasks can achieve the most encouraging results in terms of classification performance.