Tuning of vocal tract model parameters for nasals using sensitivity functions

Area of Science:

• Acoustics and Speech Science
• Signal Processing
• Computational Linguistics

Background:

• Determining vocal tract cross-sectional areas from speech signals is crucial for speech production research.
• Iterative methods using sensitivity functions are established for tuning vocal tract models to formant frequencies.

Purpose of the Study:

• To extend the application of sensitivity functions to a three-tube model for nasal sounds.
• To compare energy-based and Jacobian-based sensitivity functions for branched-tube vocal tract models.
• To evaluate iterative tuning methods for a three-tube nasal vocal tract model.

Main Methods:

• Expansion of sensitivity function concepts to a three-tube model for nasal speech analysis.
• Comparison of energy-based and Jacobian-based sensitivity functions with respect to formant frequency.
• Iterative tuning of a three-tube vocal tract model for the nasal sound /m/.
• Comparative analysis of steepest descent and Gauss-Newton methods for iterative vocal tract tuning.

Main Results:

• Energy-based and Jacobian-based sensitivity functions show negligible differences when sensitivity is calculated for formant frequency alone.
• Accurate tuning of the three-tube vocal tract model requires consideration of both polar angle and the absolute value of poles and zeros.
• The Gauss-Newton method demonstrates faster convergence than the steepest descent method, particularly with a suitable initial guess.

Conclusions:

• Sensitivity functions can be effectively applied to more complex vocal tract models, including those for nasal sounds.
• The iterative tuning process for vocal tract models benefits from considering both spectral properties (poles and zeros) and iterative solver efficiency.
• The Gauss-Newton method is a more efficient iterative solver for vocal tract model tuning compared to the steepest descent method.