Ultrasound tongue image denoising for comparison of first and second language tongue trajectories

Abstract

The main purpose of this research is to specify articulation difference between native and non-native speakers by digitizing tongue motions and analyzing the difference between utterances. Differences in tongue motion directly influence speaker’s pronunciation; therefore, it may be possible to improve non-native speaker’s efficiency of pronunciation practice with the relevant feedback and visualization. It is necessary for comparison of native and non-native speakers’ tongue motions to that end, however, normalization is absolutely necessary to remove the influence of anything except tongue motion before comparison, because every person has a unique shape and size. In our previous research, we proposed normalization methods and some speaking errors were picked up automatically from tongue trajectory in ultrasound tongue image space. However, it is necessary to improve method to extract pure tongue trajectory with accuracy. In this paper, ultrasound tongue images are separated to 5x20 or 20x5 strip block. If tongue edge lies on the block, gray scale graduation forming a shape of an arch has occurred on a line of long side of block. We proposed some shape-sensitive filter to cut off an image noise which locates excepting neighborhood of tongue edge. Through our filter, tongue trajectory extracted from ultrasound tongue image could have less noises then previous one. However, some image noises locates on neighborhood of tongue edge are emphasized and inhibited comparison of native and non-native tongue trajectories.