Three-dimensional linear modeling of tongue: Articulatory data and models

Volume images of tongue were acquired by MRI from one subject uttering a corpus representative of French allophone articulations. Supplementary images of hard palate, jaw, and hyoid bone were acquired by CT. The three-dimensional tongue surface outline was represented, for each of the 46 articulations of the corpus, by a mesh obtained by fitting a generic mesh to the set of tongue contours traced from the MR images. Jaw and hyoid bone positions were also determined. The set of the 3D coordinates of all vertices of the tongue mesh constituted the variables on which linear component analysis was applied. Six linearly independent components were found to explain 87 % of the variance of the tongue data. The associated parameters that control the linear articulator tongue model are related to jaw and hyoid positions, and to the actions of tongue muscles such as the genioglossus, the hyoglossus or the styloglossus. In addition, it was shown that the full 3D tongue surface is predictable from its 2D midsagittal contour with a mere 13.6 % increase in the overall full 3D reconstruction RMS error, which confirms quantitatively previous results. Finally, the tongue volume was found to depart by at most ±5% from its mean over the corpus, which supports the hypothesis of tongue tissue incompressibility for speech.