Local stability and long distance reliability of tongue ultrasound spline metrics

Abstract

Tongue ultrasound data arecommonly analyzed by using different metrics. In this study, we evaluate the local stability and long distance reliability of Average Nearest Neighbour Distance (Zharkova and Hewlett, 2009), Median Point-by-Point Distance (Palo, 2020), Procrustes analysis and Modified Curvature Index (Dawson et al., 2015). A metric which has good local stability should map a time-series of splines to a smooth scalar (or vector) time-series and not be very sensitive to small errors in the splining. Similarly, a metric with good long distance reliability should differentiate dissimilar tongue contours—such as different phoneme targets—reliably and without small splining errors significantly affecting the results. We test the metrics with simulated data generated from actual spline data. First, we select keyframes at phoneme target positions. Second, we vary the sampling frequency artificially by interpolating the splines between the keyframes. Third, we vary the level of splining noise by adding small perturbations to the splines at varying magnitudes and frequencies. Based on the simulation results, we will calculate the distributions of each of the metrics as conditioned by sampling frequency and noise level. The code and simulation data will be publicly available.