Automatic segmentation of vocal tract MR images

Abstract

Magnetic resonance imaging (MRI) is widely applied as a safe and reliable method in studying the hidden mechanisms of human speech production. Automatic segmentation of vocal tract shape in MRI is a challenging task due to the dynamic nature of articulation, the variability in the shape introduced by different sounds or different speakers' articulatory configurations, and the connectivity of vocal tract airway to other channels of air such as the nasal tract. A new approach for the automatic segmentation of the vocal tract shape in dynamic MR images is proposed. A method of automatic landmark tagging by recursive boundary subdivision (RBS) is applied to obtain the corresponding sets of landmarks on the vocal tract contours. The oriented active shape model (OASM) technique is adopted to recognise and delineate the shape of the vocal tract in standardised MR images. The results are presented and evaluated both qualitatively and quantitatively. We demonstrate that this is a promising approach for automatic segmentation of large databases of vocal tract images for the purposes of speech production studies.