The Science
INTERPRETING GLOTTAL FLOW DYNAMICS FOR DETECTING COVID-19 FROM VOICE
Image is an ROC.
The experiment is conducted using a 5-fold Cross validation setup. The partitioning ensured that all recordings from a selected speaker remained in the same fold — i.e. partitioning was based on speaker ids. We obtained an ROC AUC score of 0.946 with a Random Forest classifier trained by automatically filtering out some instances that exhibited lower confidence in classifying COVID vs non COVID. A threshold of 0.7 classifier confidence was used in the filtering step.
