Learning and evaluating response prediction models using parallel listener consensus