Improving vocal tract reconstruction and modeling through super-resolution volume reconstruction technique

Abstract

Magnetic resonance imaging has been widely used in speech production for vocal tract reconstruction and modeling. In order to observe detailed structures in the vocal tract, three orthogonal image stacks (sagittal, coronal, and axial) are usually acquired. Due to many constraints, each stack typically has an in-plane resolution which is much better than the out-of-plane resolution. Usually vocal tract modeling is based on just one of these three stacks. As a result, additional useful information revealed by the other two datasets is excluded in the vocal tract model. This study is to improve the vocal tract reconstruction and modeling by integrating information from all of the three stacks. To do so, a super-resolution reconstruction method recently developed to generate an isotropic image volume is used to integrate the three orthogonal stacks. Based on the ATR MRI database of vowel production, vocal tract models from MR images in high resolution, low resolution (simulated through downsampling), and super-resolution were built respectively and compared. The improvement in vocal tract modeling due to the super-resolution technique will be demonstrated on five vowels in terms of visualization and acoustic responses. [This research was supported by NIH R01 CA133015.]