An Improved Glottal Flow Model Based on Seq2Seq LSTM for Simulation of Vocal Fold Vibration

Area of Science:

• Fluid dynamics
• Acoustics
• Biomedical engineering

Background:

• Vocal fold vibration is crucial for speech production.
• Accurate fluid-structure interaction (FSI) simulations are needed for understanding vocal fold dynamics.
• Existing deep neural network (DNN) based empirical flow models (EFMs) have limitations in accuracy and efficiency.

Purpose of the Study:

• To propose an improved data-driven glottal flow model for FSI simulations of vocal fold vibration.
• To enhance the prediction performance of DNN-based EFM on accuracy and efficiency.

Main Methods:

• A Seq2Seq long short-term memory (LSTM) network infers flow rate and pressure distribution.
• Training data is derived from a generalized glottal shape library.
• The model is coupled with a solid dynamics solver for FSI simulation.

Main Results:

• The proposed model significantly reduces prediction errors compared to the EFM.
• It eliminates the need for time-consuming optimization processes.
• FSI simulations show greatly improved accuracy and efficiency.

Conclusions:

• The data-driven glottal flow model offers superior performance over previous EFMs.
• This advancement improves the simulation of vocal fold vibration for FSI analysis.
• The model enhances both the accuracy and computational efficiency of the simulations.