In self-monitoring for speech errors, some segmental speech errors are detected "early", i.e. before articulation is initiated, and others "late", i.e. after articulation has started, and others again are detected not at all. What causes these differences in detecting speech errors? We hypothesize that detection of errors that are similar to the correct target takes longer than detection of errors that are dissimilar. We also hypothesize that the time available for error detection in internal speech is limited. If this time is exceeded before an error is detected, self-monitoring gets a second chance during or after articulation. If also the time available for this second stage of error detection is exceeded, then the error remains undetected. From these hypotheses we predict that relatively more "dissimilar" than "similar" segmental errors are detected before articulation, and also that relatively more "similar" than "dissimilar" errors remain undetected. To test these hypotheses, we have assessed, in data obtained in two earlier published SLIP experiments, the relative frequencies of repaired and unrepaired segmental speech errors that differ systematically in strength of phonetic contrast with the correct target. Results show that indeed a major factor affecting the distribution of errors over "early", "late" and "undetected" is the strength of phonetic contrast between two competing response candidates.