Vocal Tract Modeling: from Signal to Structure

Tube models are very popular for the computational modeling of speech production. For nasals and nasalized vowels a minimum of two tubes is necessary to accurately model the spectral components of the speech signal. Typically, such models are estimated applying a polezero model as a first step and then estimate the tubes cross-sectional areas based on this model. Here, we introduce a method that estimates the tube areas without the necessity to estimate a pole-zero filter model. The algorithm is based on a variational Bayesian scheme under Gaussian assumptions. Probabilistic priors are used to enforce smoothness of the tubes. The method is tested on simulated data and results show that under strong smoothness prior the algorithm converges more reliably than an unconstrained method.