Toward ambulatory monitoring of vocal behavior at the physiological level using deep ensembles and Bayesian neural...

Currently, diagnosis of voice disorders is often made when patients visit the clinic, by which time speakers already experience vocal difficulties. The goal of this study was to develop a voice inversion system that predicts how speakers modulate vocal physiology from the produced voice, toward early detection of unhealthy vocal behavior. Two neural networks, a Bayesian neural network and a deep ensemble of neural networks, were developed that predict changes in vocal physiological parameters and their confidence intervals. Comparison to human data showed that both networks were able to predict meaningful differences in vocal behavior across subjects, demonstrating their potential toward ambulatory monitoring of vocal behavior at the physiological level.