Estimation of physiological vocal features from neck surface acceleration signals using probabilistic Bayesian neural networks

This study introduces a Probabilistic Bayesian Neural Network (PBNN) for estimating vocal function variables, advancing non-invasive ambulatory voice monitoring by addressing aleatoric and epistemic uncertainties in regression tasks. The PBNN estimates key physiological parameters, including subglottal pressure, vocal fold contact pressure, and thyroarytenoid and cricothyroid muscle activations, from seven aerodynamic and acoustic features. Trained on the Triangular Body-Cover Model (TBCM) of the vocal folds, the PBNN establishes a non-linear inverse mapping between inputs and outputs. These features, obtainable in ambulatory settings, enhance the practical applicability of the method. Transfer Learning integrates real voice data into the synthetic-trained network, refining subglottal pressure estimations. The PBNN generates confidence intervals that correlate prediction errors with estimated uncertainties, effectively identifying potential inaccuracies. Notably, increased uncertainty is observed at operating points with likely TBCM non-linear behaviors, such as higher subglottal pressures, indicating limitations in the selected features for capturing these effects. These findings suggest incorporating new features and additional measurements to better capture non-linear responses, paving the way for future research in the ambulatory assessment of vocal function.