Reconstructing the tongue surface from six cross-sectional contours: ultrasound data

This work presents a method for reconstructing 3D tongue surfaces during speech from ultrasound data. The method reduces the dimensionality of the tongue surface and maintains highly accurate reproduction of local deformation features. This modification is an essential step if multiplane tongue movements are to be reconstructed practically into tongue surface movements. Earlier work (Stone & Lundberg, JASA 99, 3728-3737, 1996) produced 3D reconstructions of static tongue surfaces from dense sets (60 slices) of 2D coronal tongue contours. Sparse data sets (6 slices) from within the original dense set were used to reconstruct 3D tongue surfaces, which were compared to the 60 slice surfaces. The 6 slice sets of coronal images were determined from an optimized set of midsagittal points. The reconstruction procedure was done in an identical manner to the dense data, but in 2D. Crosssectional slices of the tongue were measured at the "optimal" midsagittal points, and used to reconstruct 3D surfaces. These surfaces were compared to the dense reconstructions. Errors and reconstruction coverage were comparable to the 3D optimized sparse set, indicating this was an adequate method for calculating a sparse data set for use in reconstructing 3D surface behavior.