Article ID: e26.121
Physics-informed neural networks (PINNs) have recently attracted attention as numerical solvers for partial differential equations; however, few studies have applied them to voice production analysis. We propose a PINN framework that simultaneously learns a nonlinear vocal fold model in the time domain and a vocal tract model with frequency-dependent acoustic properties. The proposed method uses vocal tract shape parameters as input features and outputs vocal fold vibration waveforms, glottal flows, and voice waveforms at near-real-time speeds across diverse vocal tract geometries. The method is validated via comparison with reference solutions obtained by a conventional hybrid time–frequency-domain method.