Using Custom X-vectors for the Automatic Screening of COVID-19 Based on Coughing Audio Samples

A lot of effort has gone into eradicating the pandemic caused by the COVID-19 outbreak. One initiative in the efficient control of the spread of it lies in the methods for its diagnosis. Numerous techniques for screening the disease have emerged to date, which, combined with social measures, have helped to diminish the spread. Nevertheless, two years after the outbreak, the virus continues to propagate and claim victims worldwide. Therefore, there is a need for inexpensive, efficient, and real-time screening methods. In this scenario, the use of coughing samples as audio signals is a potential way to provide clinicians with an automatic tool for pre-diagnosing COVID-19 using AI techniques. This study investigates the use of coughutterances of subjects for the automatic detection of COVID-19. Relying on x-vector embeddings obtained from custom-trained deep neural network extractors on cough audio recordings, we were able to get highly competitive classification performance. Furthermore, we analyze the sensitivity of the extractors to domain dependence; and the quality of the embeddings produced in this context.