Numerical simulations for characterization of missing articulatory information in ultrasound imaging of speech

Abstract

The air–tissue interface at the tongue surface allows submental (below the chin) B-mode ultrasound images to depict the shape of the tongue in real time, useful for measuring and understanding typical and disordered tongue articulation during speech production. However, parts of the tongue are obscured; depending on an individual speaker's anatomy and tongue movement patterns, the anterior tongue tip may be shadowed by the mandible bone and the sublingual airspace. To characterize this missing information about the tongue tip in ultrasound images, midsagittal ultrasound images were numericallysimulated. Three-dimensional anatomic models were constructed using magnetic resonance imaging (MRI) data to map tongue tissue, surrounding air, and an estimated position of the mandible. Pulse-echo ultrasound images were then numerically simulated using the open-source k-Wave toolbox. Acoustic properties of the medium and an imaging array model are identified that allow simulations to replicate features in ultrasound images recorded from the same speakers as the MRI data, while maintaining computational efficiency. The potential to recover missing information on the tongue tip position from reverberation artifacts is assessed.