AutoTrace: An automatic system for tracing tongue contours

Abstract

Ultrasound imaging of the tongue is used for analyzing the articulatory features of speech sounds. In order to be able to study the movements of the tongue, the tongue surface contour has to be traced for each recorded image. In order to capture the details of the tongue’s movement during speech, the ultrasound video is generally recorded at the highest frame rate available. Detail comes at a price. The number of frames produced from even a single non-trivial experiment is often far too large to trace manually. The Arizona Phonological Imaging Lab (APIL) at the University of Arizona has developed a suite of tools to simplify the labeling and analysis of tongue contours. AutoTrace is a state-of-the-art automatic method for tracing tongue contours that is robust across speakers and languages and operates independently of frame order. The workshop will outline the software installation procedure, introduce the included tools for selecting and preparing training data, provide instructions for automated tracing, and overview a method for measuring the network’s accuracy using the Mean Sum of Distances (MSD) metric described by Li et al. (2005).