Listening for COVID: Noninvasive Detection From Cough and Breath Sounds

Abstract

Traditional procedures for diagnosing COVID-19 have mostly depended on invasive or resource-intensive technologies such as X-ray imaging, computed tomography (CT) scans, magnetic resonance imaging (MRI), and reverse transcription polymerase chain reaction (RT-PCR). Although these methods have therapeutic value, they are sometimes impractical for widespread screening, especially in settings with low resources or remote areas where access to highly skilled specialists and advanced technology is scarce. This work investigates noninvasive, AI-driven pipelines for COVID-19 detection that use cough and breath sounds as the primary inputs to address these issues. The proposed approach starts with sound acquisition and moves on to comprehensive feature extraction, focusing on image-based audio representations. The ability of various techniques, including scalograms, spectrograms, mel-spectrograms, chromagrams, wavelet spectrograms, cepstral analysis, gammatonegrams, power spectrograms, and short-time Fourier transform (STFT), to produce discriminative features that can either parallel or even enhance radiological modalities in AI-assisted systems is assessed.