A Deep Learning-Based Generalized Empirical Flow Model of Glottal Flow During Normal Phonation

Area of Science:

• Computational Fluid Dynamics (CFD)
• Bioacoustics
• Deep Learning

Background:

• Accurate modeling of glottal flow is crucial for understanding voice production.
• Traditional methods like Navier-Stokes (N-S) solutions are computationally expensive.
• Existing empirical models often lack generalizability and accuracy.

Purpose of the Study:

• To develop a fast and accurate deep learning-based generalized empirical flow model (EFM) for glottal flow prediction.
• To integrate the EFM with fluid-structure interaction (FSI) simulations for phonation.
• To evaluate the EFM's predictive performance against high-fidelity N-S solutions.

Main Methods:

• Generated a glottal shape library using a universal kinematics equation (UKE).
• Obtained ground truth flow rate and pressure data via high-fidelity N-S solutions.
• Trained a deep neural network (DNN) to create an empirical mapping for the EFM, coupled with a finite element method (FEM) solver for FSI.

Main Results:

• The DNN-based EFM demonstrated accurate prediction of glottal flow and pressure distributions.
• Integration with FEM-based FSI simulations showed good performance.
• The EFM achieved significant improvements in computational efficiency compared to N-S solutions.

Conclusions:

• The proposed deep learning-based EFM offers a computationally efficient and accurate approach for simulating glottal flow.
• This model holds potential for advancing research in voice production and related clinical applications.
• The EFM provides a viable alternative to traditional high-fidelity simulations for phonation analysis.