Unmasking COVID-19 symptoms from Speech with Masked Autoencoders

In the aftermath of the COVID-19 pandemic, the need for swift, reliable, and scalable methods for COVID-19 like symptom detection has come into sharper focus. In response, we introduce Audio Masked Harmonic Autoencoder (AudioMHA), a novel self-supervised deep learning model designed to identify respiratory symptoms associated with COVID-19 and similar illnesses from audio signals, particularly cough sounds. Our model harnesses the strength of the transformer architecture, integrating an innovative Harmonic Attention-based decoder and a two-stage training procedure. Initially pre-trained on the rich and diverse AudioSet corpus, the model is then fine-tuned on the specific COUGHVID dataset, enabling the acquisition of robust, generalizable audio representations before honing in on detecting COVID-19-like cough signatures. Delivering an impressive F1 score of 86%, our model outperforms previous state-of-the-art baselines by a significant margin of 10% . This pioneering, non-invasive, and potentially scalable approach marks a significant stride towards enhancing preliminary symptom detection methods for respiratory illnesses. Our findings highlight the potential of sophisticated machine learning techniques in future healthcare applications, paving the way for prospective research in audio-based symptom detection.