Two experiments on words isolated from recorded dictations show that the effect of repetition on intelligibility depends on the discourse roles of the tokens being compared. When the later token adds no new information to the discourse, as in self-corrections and other coreferential repetitions, intelligibility falls with repetition. When the later token introduces a new discourse entity, intelligibility may rise with repetition. The results are discussed with reference to the effects of the speaker's experience in machine dictation and to the consequences f0 automatic speech recognition.