Machine learning surrogate forward models for biomechanical laryngeal control.

Area of Science:

• Biomechanics
• Computational Auditory Neuroscience
• Machine Learning

Background:

• Accurate laryngeal biomechanical models are crucial for understanding voice production.
• Traditional ordinary differential equation (ODE) models face computational challenges, hindering real-time applications.
• Machine learning (ML) offers potential for faster, more stable voice production modeling.

Purpose of the Study:

• To evaluate ML regressors (Random Forest, Neural Networks, Polynomial Regression) as surrogate forward models for laryngeal motor control.
• To compare the performance of ML surrogates against traditional biomechanical models in terms of speed and stability.
• To assess the impact of training data size on ML surrogate model accuracy.

Main Methods:

• Generated training data using two biomechanical vocal fold models (extended body-cover and triangular body-cover).
• Trained and evaluated Random Forest (RF), Multilayer Perceptron Neural Networks (NN), and Polynomial Regression (PR) as ML surrogate models.
• Measured execution times and assessed control signal smoothness for ML surrogates.

Main Results:

• ML surrogates drastically reduced execution times from seconds to milliseconds (e.g., 2 ms for PR).
• Random Forest (RF) achieved the highest accuracy, while Neural Networks (NN) and Polynomial Regression (PR) provided smoother control signals.
• Model accuracy degraded significantly below approximately 1,000 training samples.

Conclusions:

• ML surrogates are efficient and adaptable alternatives to direct numerical simulations for laryngeal biomechanical modeling.
• These models enable stable, real-time tracking via inverse Jacobian control.
• Findings support the use of ML surrogates for future subject-specific voice modeling, particularly through transfer learning in data-limited settings.