Machine learning surrogate forward models for biomechanical laryngeal control

Abstract

Accurate modeling of laryngeal motor control is key to understanding typical and disordered voice production. However, traditional biomechanical plant models based on ordinary differential equations (ODEs) often involve high computational costs and numerical instabilities, limiting their use in real-time closed-loop control frameworks. This study evaluates feature-driven machine learning (ML) regressors, specifically Random Forest (RF), Multilayer Perceptron Neural Networks (NN), and Polynomial Regression (PR), as surrogate forward models mapping laryngeal motor inputs to fundamental frequency and sound pressure level. Training data were generated with two biomechanical vocal fold models: the extended body-cover and the triangular body-cover. Results demonstrate that ML surrogates reduce execution times from seconds to milliseconds (e.g., 2 ms for PR), enabling stable real-time tracking via inverse Jacobian control. While RF provides the highest accuracy, NN and PR offer smoother control signals and smaller memory footprints. A practical performance threshold was identified near N = 1,000 training samples, below which accuracy degraded substantially when models were trained from scratch. These findings support ML surrogates as efficient and adaptable alternatives to direct numerical simulation, providing a foundation for future subject-specific modeling through transfer learning in data-limited clinical scenarios.