Silent speech with ultrasound

We trained a model to predict speech from ultrasound video of the tongue, no speaking required. It reads open text, generalizes across people, and reaches a 15.6% word error rate on open-vocabulary speech.