Errors in determining vocal tract shape from the acoustic signal

This paper presents the results of an experimental study on the errors associated with determining the shape of a vocal tract from the acoustic output signal. A vocal tract model was constructed from plexiglas sections of constant thickness, each having a hole of a specified area, that closely approximated a lossless autoregressive (AR) system. A small wide-band speaker acted as the excitation source. Linear prediction procedures were applied to the observed signals to estimate the shape of the tube. These values were then compared to the actual values to determine the error. The results indicate the experimental and signal processing conditions required for the estimates of the vocal tract shape to approximate the actual shape.