Cough Sound Classification Based on Similarity Metrics

Abstract

Cough and respiratory sound processing can assist in the early diagnosis of infections such as Covid-19. Even asymptomatic Covid-19 patients can be diagnosed early enough if appropriate speech modeling and signal-processing is applied. Covid-19 affects various speech subsystems that are involved in respiration, phonation and articulation. Based on a symptom tracking platform that was recently presented by the authors (Coronario), we focus on the sound processing subsystem that is capable of classifying cough or respiratory sounds in multiple categories. Specifically, we attempt to classify a cough sound file in one of the following 5 categories: male dry or productive, female dry or productive and child's cough. The classification is performed using Pearson Correlation Similarity, in frequency domain. Several alternative methods that employ averaging and Principal Component Analysis have been tested to estimate their recall and precision/accuracy metrics. The average precision/accuracy achieved is about 75% and 88%, respectively. The sound processing platform used is extensible allowing researches to experiment with several different classification methods applied on the anonymized data exchanged during symptom tracking. © 2021 IEEE.