Researchers have developed a machine learning model that enables silent speech by predicting words from ultrasound video of the tongue without requiring the user to vocalize. The system generalizes across different individuals to decode open-vocabulary text, achieving a 15.6% word error rate.