Estimation of Physiological Vocal Features from Neck Surface Acceleration Signals Using Probabilistic Bayesian Neural...

Area of Science:

• Bioacoustics and Speech Science
• Computational Neuroscience
• Biomedical Engineering

Background:

• Accurate estimation of vocal function parameters is crucial for diagnosing voice disorders.
• Non-invasive ambulatory voice monitoring requires robust methods to handle real-world data variability.
• Existing methods often struggle to quantify uncertainty in physiological parameter estimations.

Purpose of the Study:

• To develop and validate a Probabilistic Bayesian Neural Network (PBNN) for estimating key vocal function variables.
• To enhance non-invasive ambulatory voice monitoring by addressing aleatoric and epistemic uncertainties.
• To refine estimations using transfer learning with real voice data.

Main Methods:

• A Probabilistic Bayesian Neural Network (PBNN) was developed for inverse mapping of aerodynamic and acoustic features to physiological parameters.
• The PBNN was trained using the Triangular Body-Cover Model (TBCM) of vocal folds with synthetic data.
• Transfer learning was employed to integrate real voice data, refining subglottal pressure estimations.

Main Results:

• The PBNN successfully estimated subglottal pressure, vocal fold contact pressure, and muscle activations.
• Confidence intervals from the PBNN correlated with prediction errors, indicating effective uncertainty quantification.
• Increased uncertainty was observed at higher subglottal pressures, suggesting limitations in capturing non-linear vocal fold dynamics.

Conclusions:

• The PBNN offers a promising approach for non-invasive, uncertainty-aware estimation of vocal function parameters.
• The method enhances ambulatory voice monitoring by providing reliable confidence intervals for estimations.
• Future research should explore additional features to better capture non-linear vocal fold behaviors.