Speech signals as biomarkers: using glottal features for non-invasive COVID-19 testing

The COVID-19 pandemic was the most significant global health crisis in recent history, with lasting impacts on societies worldwide. Current screening methods are invasive, slow, frequently inaccurate, and limited in capacity. To overcome these limitations researchers used conventional features extracted from speech signals. In the proposed methodology, the change in vibratory pattern due to COVID-19 is captured by extracting glottal features from glottal signal acquired by inverse filtering. Various machine learning models like naïve Bayes, ADABOOST, gradient boost, decision tree, support vector machine (SVM), stochastic gradient descent (SGD), K-nearest neighbours, and random forest (RF) are used in this study. It is observed that the gradient boost gives the maximum classification accuracy (99.97%) for COVID-19 detection and SGD gives the maximum classification accuracy (99.76%) for severity grading. It is also observed that the time instants glottal features outperform frequency domain and model-based features. The Coswara database is used for this study.