The purpose of this thesis was to explore experimentally the structures and the processes that are necessary for the production of speech. A number of hypotheses were tested by asking the subjects to perform secondary tasks while they talked. Experiments 1 and 2 investigated whether speakers must pause in order to plan. The subjects read one sentence aloud and were given the topic for a subsequent sentence that they had to make up. The results showed that the subjects could use the prior information without a reduction in reading speed. In Experiment 3 two judges rated graphs produced from monologues for the presence of a molar rhythm of phonation and silence. The ratings showed that such rhythms are not typical of speech. Furthermore, the judges could not distinguish these graphs from a matched set of computergenerated random walks. Experiments 4,5 and 6 examined the deployment of stored information during speech. The subjects produced sentences faster from related rather than unrelated noun pairs, and with a digit preload rather than without. The findings suggested that speakers must always do semantic work, though the results from Experiment 7 showed that subjects were more fluent for topics in which past events were recalled. The question of which units occur during speech planning was investigated in Experiments 8 and 9 by asking the subjects to talk while they were tracking. The pattern of tracking errors suggested that speakers translate their ideas into propositional units. However, if a propositional function or argument is not expressed, the referents in working memory are checked in order to determine that the verb role requirements have been satisfied. These results suggest that there is no level of psychological representation equivalent to deep structure. Finally, the results from these studies are incorporated into a model of speech production.