Detection of COVID-19 in Respiratory Sounds using End-to-End Deep Audio Embeddings

• DOI: 10.22533/at.ed.1592332230061

• Palavras-chave: audio classification, cough sounds, COVID-19 detection, deep learning, respiratory sounds, transfer learning

• Keywords: audio classification, cough sounds, COVID-19 detection, deep learning, respiratory sounds, transfer learning

• Abstract:

  Due to the COVID-19 worldwide pandemic situation, automatic audio classification research has been of interest for analysis of respiratory sounds. Several deep learning approaches have shown promising performance for distinguishing COVID-19 in respiratory cycles. In this work we explored the usage of transfer learning from a pre-trained end-to-end deep-learning based audio embeddings generator named AemResNet, applied to the classification of respiration and coughing sounds into healthy or COVID-19. We experimented with the publicly available large-scale Cambridge Crowdsourced dataset of respiratory sounds collected to aid diagnosis of COVID-19. Our presented work focuses into 3 experimental tasks: 1) detection of COVID-19 from a combination of breath and cough sounds, 2) detection of COVID-19 from breath sounds only, and 3) detection of COVID-19 from cough sounds only. The experimental results obtained over this respiratory dataset show that a pre-trained audio embedding generator achieves competitive performance compared to the recent published state-of-the-art.

• Número de páginas: 4