Parameter sharing and its effects for structural speech recognition