Regularising an adaptation algorithm for tongue shape models

Realistic data-driven models of the tongue shape can be obtained by learning a nonlinear mapping from tongue landmarks to full contours, trained on a dataset of thousands of contours. Semiautomatic contour extraction from ultrasound takes a lot of time and effort from an expert, so practically it is preferable to adapt a reference model given just a few contours from the new speaker. However, adaptation with very few contours is unreliable and prone to overfitting. We study several forms of regularisation to constrain the adaptation, and determine the optimal amount of regularisation by leave-one-out cross-validation. Our results show that good accuracy models can be found reliably with no user intervention.