Discrete representations in neural models of spoken language