Evaluation of automatically generated tongue surfaces from three-dimensional/four-dimensional ultrasound recordings of sustained vowels

Three-dimensional/four-dimensional (3-D/4-D) ultrasound is capable of imaging almost the entire tongue surface during the production of speech sounds, including sustained vowels. Manual tongue surface segmentation is slow and laborious, but typically yields good results. Automatic tongue surface segmentation is much faster and objective but can result in errors. In this presentation, we quantify the errors in a set of tongue surfaces that were automatically segmented using the previously published 3D-SLURP algorithm implemented in the WASL toolbox for MATLAB. The ultrasound recordings were of 20 children and 40 adults producing the sustained vowel [o:] with and without a narrow straw semi-occlusion in the mouth. Twenty of the 40 adults produced phonation at threshold phonation. Automatic and manual error detection and quantification methods will be compared to evaluate the accuracy of the automated algorithm when applied to the sustained vowel [o:]. [Work supported in part by NSF.]