Translating pellet positions into constriction features

Abstract

Tongue constriction features can be estimated from sagittal x-ray pictures of the tongue surface and vocal tract wall. However, such records cannot be obtained in quantity, making them unsuitable for testing models such as Stevens's quantal theory. The x-ray microbeam allows larger data sets, but records flesh points rather than surfaces. This paper presents an algorithm for relating the two representations. The vocal tract wall is estimated from whole-head scans and a palate trace. Pellet positions are then “warped” into a Cartesian space where location along the tract and distance from it are the x and y values. The algorithm has been applied in a replication of Perkell and Nelson's test of quantal theory using principal component analysis. Quantal theory predicts that the pellet closest to the constriction site will show least variability, and that the most precision will be in the dimension perpendicular to the vocal tract wall for “quantal” vowels such as /i/. In the warped space, then, the principal component of variation for the relevant pellet for these vowels should be parallel to the x axis. This prediction is borne out. [Work supported by the NSF.]