This study examined the effects of task demand on the frequency of speech errors. Subjects spoke continuously in three levels of task demand: low, moderate, high. Task demand was manipulated through the use of an additional task. The frequency of various speech errors increased with increased task demands. Of particular interest were sound errors that are lexically legitimate (word) and anomalous (non-word). With low task demands, more word errors were produced than non-word errors. Under moderate and high task demands, more non-word errors were produced than word errors. More non-word errors were produced at moderate and high task demands than at low task demands while the number of word errors remained unchanged. These results are consistent with earlier studies on sound error elicitation and lend support for an editing process that detects and corrects errors prior to articulation. Accordingly, under low task demands non-word errors are rarely articulated because the editor, through access to the lexicon, easily detects and corrects non-word errors but not word errors. The results also suggest that the editor requires cognitive resources. As task demands increase, resources become unavailable to the editor allowing the non-word errors to slip past the editor as do word errors.