Accuracy assessments of hand and automatic measurements of ultrasound images of the tongue

Ultrasound measurements of the tongue have become increasingly useful in phonetic research, but the consistency of hand measurements and the accuracy of automatic measures have yet to be extensively evaluated. Here, we examine three automatic methods (EdgeTrak, active contours (snakes) and SLURP) against hand measurements of two datasets made by three researchers. Discrepancies (errors) were calculated as the shortest distance between two contours. Results for adult speakers indicate that the three researchers were consistent with themselves and each other (mean errors < 1 mm) but higher for child data (errors < 5 mm). Automatic methods were initially consistent with hand measurements, but snakes and EdgeTrak were increasingly less consistent at later points. The results indicate that the SLURP method can be used for automatic extraction of full tongue shapes from ultrasound images. Intermittent hand correction of automatic procedures is also recommended. Future work will include more systems and more speakers.