Looking to Listen at the Cocktail Party: A Speaker-Independent Audio-Visual Model for Speech Separation.