Early work showed that semantic, but not phonological, working memory (WM) supports the ability to produce multiword utterances.^1,2 Recent results from narrative production at the acute stage of stroke corroborated these findings, with semantic, but not phonological WM, predicting narrative measures of sentence elaboration.³  However, phonological WM was found to have a positive relation with speech rate (words per minute), and a negative relationship with proportion pronouns relative to nouns.^3,4 Two hypotheses might explain these relationships: 1) slower, more error-prone phonological retrieval of single words leads to slower speech rate, reduced rehearsal in WM tasks, and a preponderance of pronouns (which are easy to retrieve),³ and 2) the existence of separate input and output phonological WM buffers,^4,5,6 where the output buffer supports fluent speech, phonological WM, and the maintenance of longer words. Follow-up analyses supported the phonological retrieval hypothesis; however, input and output WM capacities could not be distinguished in the acute sample.⁵ The current study evaluates these two hypotheses for individuals with chronic aphasia, where measures of input and output phonological WM were available.