Machine learning surrogate forward models for biomechanical laryngeal control

Area of Science:

• Computational biophysics
• Voice production mechanisms
• Machine learning applications

Background:

• Accurate laryngeal motor control modeling is vital for voice production research.
• Traditional biomechanical models are computationally intensive for real-time applications.
• Need for efficient alternatives to direct numerical simulation.

Purpose of the Study:

• Evaluate machine learning (ML) regressors as surrogate forward models for laryngeal motor control.
• Compare random forest (RF), neural networks (NN), and polynomial regression (PR) performance.
• Assess ML surrogates for real-time voice tracking and subject-specific modeling.

Main Methods:

• Generated training data using extended and triangular body-cover vocal fold models.
• Trained RF, NN, and PR models to map laryngeal inputs to fundamental frequency and sound pressure level.
• Evaluated inference speed, accuracy, control signal smoothness, and memory footprint.

Main Results:

• ML surrogates achieved millisecond-scale inference (e.g., 2ms for PR), enabling real-time tracking.
• RF offered highest accuracy; NN and PR provided smoother control and smaller memory usage.
• Performance degraded significantly below ~1000 training samples.

Conclusions:

• ML surrogates are efficient, adaptable alternatives to direct numerical simulation for laryngeal modeling.
• These models facilitate real-time inverse Jacobian control and hold potential for subject-specific modeling via transfer learning.
• Identified a practical training data threshold for reliable ML model performance.