Self-intersection penalty for predicted vocal-tract shapes

Develop and incorporate a penalty for self-intersection into elastic kernel ridge regression so that predicted vocal-tract contours satisfy physical non-self-intersection constraints.

Background

In the temporal smoothing application, the fitted model does not impose physical constraints on predicted vocal-tract shapes. Consequently, reconstructed boundaries can intersect in frames where the tongue touches the palate, producing configurations that are not physically plausible. The paper identifies adding a self-intersection penalty as a concrete future problem, although the issue occurs in only a small proportion of frames.