Existing large-scale articulatory databases describe the tongue shape through the 2D positions of 3–4 fixed landmarks on the tongue surface. The ability to reconstruct the full tongue contour from these landmarks would increase the utility of these databases in speech research. We give an algorithm to adapt a predictive model of the tongue contour, that has been learned using ultrasound data for a given speaker, to a new speaker for which only landmark coordinates are given. We show realistic reconstructions of the full tongue contour in the MOCHA and XRMB databases.
Chao Qin, Miguel Á. Carreira-Perpiñ&