Speech error data have been used to argue for the psychological reality of distinctive features and phonemes as well as the hierarchical ordering levels of processing for speech production. The models of production that have emerged from analysis of these data are nearly unanimous in characterizing (implicitly or explicitly) the motor output level as entirely governed by prior selection and processing of larger units, especially the phoneme. This study reports on the laboratory elicitation of sublexical speech errors by means of tongue twisters. Simultaneous audio and electromyographic recordings were analyzed. Where possible, single-motor unit discrimination was carried out to preclude the possibility of signal contamination by activation of adjacent musculature. The results indicate that traditional methods of data collection on which most speech error corpora are based are inadequate. Production models based on these corpora are not supported by the electromyographic data and must accordingly be revised.