Levelt (1983) proposed that the reason why speakers often do not interrupt an erroneous word before self-repairing is a lack of trouble detection before the end of that word. However, this explanation does not apply to merely inappropriate words. According to Levelt, this latter kind of word is completed for pragmatic reasons. In the present paper, a new corpus containing 1225 repairs is analyzed. From Levelt's theoretical framework, it was predicted that for erroneous words the longer the reparandum the higher the amount of word interruptions. Another prediction was that this decrease across word length should be slighter for nonerroneous words than for erroneous ones. Both predictions were confirmed. Results were consistent with Levelt's hypothesis and especially with the idea that erroneous word completion is not a real exception to the main interruption rule in speech self-repairing.