Speaker
Description
The human tongue is a muscular hydrostat whose intricate, three-dimensional muscle activity drives nearly every aspect of articulate speech. In this talk, I present an integrated computational framework for studying tongue biomechanics during speech using tagged and diffusion MRI. First, I introduce a sparse non-negative matrix factorization approach that decomposes voxel-level motion into functional units—data-driven groups of cohesive local muscle regions that compress, expand, and move together during protrusion and simple speech tasks—yielding subject-specific building blocks of lingual behavior. Second, I describe how Granger causality analysis applied to fiber-aligned strain time-series reveals the sequential, predictive interactions among individual tongue muscles and among the functional units themselves, illuminating how coordinated motion unfolds over time during utterances. Third, I show how these internal motion representations can be linked back to the acoustic signal through a plastic transformer that synthesizes speech audio directly from tagged-MRI weighting maps, closing the loop from anatomy to motion to sound. Together, these complementary analyses move from anatomy to motion to acoustics, offering speech scientists a quantitative, subject-specific window into normal articulation and a principled foundation for understanding, predicting, and ultimately rehabilitating disordered speech in clinical populations such as glossectomy, ALS, and motor speech disorders.