A Model for Speech Unit Duration