Did they try to predict a person's thoughts ? That would be more compelling to me than 500ms delay between the model prediction and the spoken word.