Six normal volunteers were studied with positron emission tomography to identify the cortical neural networks that participate in the processing of single words. Activity-related changes in regional cerebral blood flow were measured consecutively on 6 occasions in each subject, 2 while the subject was at rest and 4 while single word language tasks were being performed. The data from each subject were standardized for brain shape and size, reconstructed parallel to the intercommissural line, normalized for global flow differences, and then averaged for each activation condition across the 6 subjects. Significant areas of change in rCBF (P less than 0.05, with appropriate Bonferroni corrections) between task and rest conditions were displayed with reference to the coordinates of a standard neuroanatomical atlas. We have demonstrated that categorical judgements on heard pairs of real words activate neural networks along both superior temporal gyri, but with an anatomical distribution no different from that seen when the subjects listened to nonwords: the tasks would appear to be very different in cognitive demands but not in terms of the distribution of activation. However, during a verb generation task that involved thinking of verbs appropriate to heard nouns presented at a slow rate, the only temporal region activated was the left posterior superior temporal association cortex (Wernicke's area). Further analysis showed that whereas activation in other superior temporal regions, both left and right, correlated with rates of word presentation during the 4 tasks, there was no such correlation in Wernicke's area; evidence that this site is responsible for more than early acoustic processing. During verb generation there was also activation of left premotor and prefrontal cortex (including Broca's area and the supplementary motor area). The supplementary motor area is thought to be involved in the motor planning of speech. The subjects did not vocalize during the task, and therefore it would appear that the act of retrieving words from semantic memory activates networks concerned with the production of speech sounds. We conclude that single word comprehension and retrieval activate very different distributed regions of cerebral cortex, with Wernicke's area the only region engaged by both processes and with participation during silent word generation of networks involved in vocalization.