Experimental comparisons ideally should compare the shape of the tongue under different test conditions, looking for reliable differences. However, tongue outlines are inherently curvilinear, leaving standard regression analyses inappropriate. One potential resolution is to apply smooth-spline regression, which can establish a curvilinear function of best bit and yield confidence intervals. For example, the graph below shows smooth spline regressions of a subject's tongue under different coarticulatory conditions.
A localized effect is one in which part of the tongue shape remains stable (especially near the active point of contact) while the remainder varies with the surrounding phoneme. For example, /t/ may maintain a stable laminal shape regardless of nearby vowels, while the dorsum may vary widely. In contrast, in distributed coarticulation, even the point of contact may differ, and so the entire shape of the tongue differs across coarticulatory environments. Crucially, these differences are best detected by tracking the outline of the tongue, rather than with point-tracking methodologies.
One methodological issue in the ultrasound data collection paradigm is the difficulty in combining data from the same subject collected at different times. Thus, in one current project, I am using optimization of fit to match tongue curves from different sessions. In this procedure, corresponding test items from different sessions are used for calibration; the parameters to transform one curve to its best fit can then be applied to all items from the same session. For example, we can rotate and translate the /ae/ of had from one subject's session 2 so that it best fits the same subect's /ae/ from session 1. We can then apply the same rotation and translation parameters to all other items from the subject's session 2 - ideally, they should fit their corresponding session 1 items just as well. The figures below provide an illustration.