Vocal Tract Morphology Using Real-Time Magnetic Resonance Imaging

Abstract

Real-time Magnetic Resonance Imaging (rtMRI) leads to the dynamic observation of hidden processes of articulation in an unprecedented way. The non-invasive image acquisition nature of MRI combined with enhanced anatomical discrimination made rtMRI the reference in capturing vocal tract configurations during speech production. However, this development also unveiled challenges, such as the shape extraction and analysis of the vocal tract contours automatically. This work describes automated techniques for the segmentation of the vocal tract and identification of articulatory structures using rtMRI. The identification of these structures is vital for modeling articulatory synthesis. The methodology is based on level set methods to outline the vocal tract shape. Changes in the vocal tract shape and its structures were investigated for different corpora in order to bind the expression of phonemes and the behavior of the anatomical shapes. These shapes were labeled from basal form invariants, whose final evolution yielded the classification of regions of interest. The methodology resulting from this work may be employed in accent-suppression systems, speech production for laryngectomized patients, and therapetic techniques for children suffering from speech apraxia.