Direct posterior confidence for out-of-vocabulary spoken term detection