DIFFERENTIATING NORMAL VS MISARTICULATED TONGUE TRAJECTORIES FROM ULTRASOUND FOR FAST AUTOMATIC ARTICULATORY BIOFEEDBACK

Ultrasound images of the tongue surface have been increasingly used for immediate visual feedback while a speaker is working to change pronunciation. However, the tongue surface contour is complex and changes rapidly, meaning that speakers find it difficult to compare undesired vs. desired tongue shapes under realistic speaking conditions. For biofeedback in speech therapy, it is important to identify deviation from the desired tongue trajectory in real time. We present results from a preliminary study characterizing differences between misarticulated and normally articulated (“accurate”) American English /ɑɹ/ using an efficient, automatic tongue surface tracking method that separately characterizes motion of the blade, dorsum and root. Results from principal component analysis of trajectory data from child speakers show that misarticulated trajectories are distinctly different from accurate trajectories. A statistical model for classifying accurate vs. misarticulated productions is discussed.