A Fast Robust 1D Flow Model for a Self-Oscillating Coupled 2D FEM Vocal Fold Simulation

Arvind Vasudevan, Victor Zappi, Peter Anderson, Sidney FelsInterspeech 2017

TL;DR

This paper introduces a novel, computationally efficient 1D fluid model for simulating vocal fold dynamics, coupled with a 2D FEM structural model. This approach achieves a balance between the speed of simplified models and the accuracy of complex continuum models, enabling faster and higher-quality articulatory speech synthesis.

What Problem Does This Paper Address?

Articulatory speech synthesis aims to generate speech by simulating the physical movements of the vocal tract. A crucial component of this is accurately modeling the vocal folds, which act as the sound source. Existing models fall into two camps: lumped-element models are fast but less accurate, while continuum models (like 2D Navier-Stokes) are more accurate but computationally prohibitive for real-time synthesis. This paper addresses the need for a vocal fold model that bridges this gap, offering robust and fast simulations suitable for articulatory speech synthesis pipelines. Specifically, it tackles the challenge of accurately capturing unsteady flow phenomena, glottal closure, and complex geometries within a computationally feasible framework.

Methods

The core innovation is a 1D unsteady fluid model that is "loosely coupled" with a 2D Finite Element Method (FEM) structural model of the vocal folds.

  • Fluid Model: Instead of computationally expensive 2D Navier-Stokes equations, the authors adapt a 1D fluid model (based on Anderson et al.) that handles irregular geometries, boundary conditions, and unsteady flow states. A key contribution is a fast, decoupled numerical scheme for solving the flow equations, avoiding the need for Jacobian calculations, which significantly speeds up computation, especially for pressure-driven flows. They also introduce a method to convert pressure-pressure boundary conditions to velocity-pressure conditions to leverage this faster solver.
  • Structural Model: A 2D FEM model, based on previous work by Alipour et al., is used for the vocal fold structure. It employs linear-elasticity for computational efficiency and is divided into body, cover, and ligament regions.
  • Coupling: The fluid and structural models are "loosely coupled." The structural model provides the cross-sectional area function (A(x,t)) to the fluid model, and the fluid model's aerodynamic pressure is used as a forcing vector for the structural model. A crucial aspect of the coupling is an "area-warping" method to handle glottal closure and prevent numerical instabilities when the vocal folds collide.
  • Acoustic Simulation: The output of the vocal fold model (glottal flow $U_g$) is fed into a 2D FDTD vocal tract solver and a 1D wave-reflection analog model for the trachea, forming a complete articulatory synthesis pipeline.

Datasets

Not discussed in the paper. The authors refer to existing literature for validation and use data from magnetic resonance imaging (MRI) of vocal tract shapes (Story et al. [25]) to create the vocal tract model for the vowel /a/.

Key Insights and Ideas

  • The 1D Fluid Model Advantage: A 1D fluid model, when carefully designed, can capture essential unsteady flow phenomena (like separation and jet regimes) that simpler Bernoulli models miss, while being significantly faster than 2D Navier-Stokes.
  • Decoupled Solver Efficiency: The proposed decoupled numerical scheme for the 1D fluid model offers a substantial speedup (45-50x) compared to coupled solutions, making real-time simulation more feasible.
  • Robustness to Closure and Collision: The area-warping mechanism is critical for handling glottal closure and collisions robustly, a common challenge in vocal fold modeling.
  • Bridging the Accuracy-Speed Gap: The paper demonstrates a successful strategy to achieve a better balance between simulation speed and physical realism, which is vital for articulatory speech synthesis.

Key Results

  • Speedup: The decoupled 1D solver was 45-50 times faster than a coupled 1D solver and more than an order of magnitude faster than a 2D Navier-Stokes solver.
  • Phonation Characteristics: The self-oscillating model produced a fundamental frequency of approximately 146 Hz, with peak flow rate $U_g$ of 473 mL/s, mean flow rate of 241 mL/s, and maximum glottal width of 1.67 mm.
  • Pressure and Velocity Distributions: The model generated realistic spatial pressure and velocity distributions, showing good agreement with literature data from more complex models (e.g., Alipour et al. [21]). Peak velocities corresponded to the minimum glottal diameter, and pressure recovery was observed.
  • Stability: The model remained stable during collisions and complex glottal shapes.

Limitations and Future Work

  • Limitations:
    • The model uses a linear-elastic assumption for the vocal fold structure, which might not capture all nonlinear material behaviors.
  • The 1D fluid model is an approximation and may not capture all the nuances of 3D flow, particularly turbulence in complex geometries.
  • The paper notes that the sharp tail-off in velocity in the divergent glottis might be due to the simplified x transition function and the absence of false vocal folds, unlike some more complex models.
  • A sudden pressure spike at the glottal exit in the neutral vocal fold shape is attributed to numerical imprecision from the x transition function.
  • Future Work:
    • Further validation is needed by examining glottal flow ($U_g$), phonation onset frequency, and the overall sound quality of the synthesized speech.
  • Optimization of the current implementation is suggested, including speeding up the matrix solver and parallelizing the fluid and trachea models.
  • Exploring more advanced material models for the vocal fold structure.

Summary updated

BibTeX
@inproceedings{vasudevan2017fast,
  author = {Arvind Vasudevan and Victor Zappi and Peter Anderson and Sidney Fels},
  title = {{A Fast Robust 1D Flow Model for a Self-Oscillating Coupled 2D FEM Vocal Fold Simulation}},
  booktitle = {Interspeech},
  year = {2017},
}

Loading related content...

Figure