Implementing artificial intelligence models for automatic vocal tract image segmentation

Modeling articulatory representations is critical to the scientific study of speech production and has been addressed by quantifying dynamical movements of the vocal tract. However, modeling dynamic articulatory data has proven time-consuming and computationally taxing. For example, real-time imaging data of the vocal tract has been assessed using contour-tracking methods for image segmentation that require manually creating vocal tract templates and human supervision to correct image segmentation (e.g., Bresch & Narayanan 2009). Recently, new work has introduced multimodal approaches to model articulatory data (Jain et al., 2024), but these approaches still require manual initiation/correction. We combine previous segmentation methods such as the Chan-Vese Segmentation algorithm (Getreuer 2012) with AI models like the Segment Anything Model (SAM) (Kirillov et al. 2023) to test means for optimizing image segmentation of vocal tract articulators in real-time magnetic resonance imaging (rtMRI) data. We deploy non-linear filtering and parameter optimization so as to prioritize desirable image processing features, such as smoothing the image while sharpening the edges. Processing these optimized images into SAM allows for segmentation of critical vocal tract articulators and the airway. In sum, the goal of this work is to automatically segment without manual intervention moving articulators in rtMRI speech production data. [Work supported by NSF.]