Models of perceptual systems customarily characterize their maximally efficient operation in optimal circumstances. Another engineering consideration - graceful failure - is usually ignored. Three experiments on spontaneous speech show that on-line speech recognition fails gracefully by making us deaf to the words in reparanda. the items which must be expunged to restore disfluent utterances to fluency. Experiment 1 uses word-level gating of fluent and disfluent utterances to show that disfluencies principally disrupt normal late recognition (Bard, Shillcock & Altmann, 1988) of words in reparanda. Experiment 2 shows that in more natural listening conditions, attention to continuing material and additional effects of repetition deafness (Miller & Mackay, 1996) make recall of the same words even more unlikely. Experiment 3 shows that the results are not attributable to the clarity of the lost words. Finally the relationships among late recognition and various kinds of disfluency deafness are discussed.