This paper investigates the effects of lexical stress pattern on speech production by aphasic subjects. The subjects are more accurate at producing words with first syllable stress than those with stress on the second syllable in both word and non-word repetition and picture naming. The subjects' accuracy in repetition is unaffected by the stress pattern of a preceding word in a phrase, but strongly affected by the phrase's syllable length. With three-syllable words, subjects are most accurate in both repetition and naming with words whose primary stress falls on either the first or the third syllable. Explanations of these results in terms of the vulnerability of initial unstressed syllables, or simple frequency-dependent stress assignment strategies, are rejected. Instead it is argued that an utterance's segmental and metrical structure are computed in parallel, but the computation requires resources that are limited in subjects who make phonological errors in speech production.