Previous work has supported a critical role for semantic, but not phonological, working memory (WM) in the ability to produce multiword utterances. In narrative production of patients at the acute stage of stroke it was shown that correlations between semantic, but not phonological WM, and narrative measures of elaboration remained when single word production abilities were included in the model. However, the acute stroke results also revealed relations between phonological WM and narrative production – specifically, a positive relationship between digit matching span (a phonological WM measure) and words per minute, and a negative relationship between digit matching span and proportion pronouns relative to nouns. Two hypotheses have been compared to explain these relationships. The first is that both digit matching span and these narrative measures relate to the speed and ease of phonological retrieval. The second is that there are separate input and output phonological WM buffers, and the output buffer plays a key role in fluent speech and phonological WM. In a sample of participants at the chronic stage of stroke two approaches, a case series and a case study approach, were used to address these two hypotheses. The case series approach showed evidence that single word retrieval abilities predicted words per minute above changes in output phonological WM, supporting the phonological retrieval hypothesis. However, the results for proportion pronouns differed between the acute and chronic stages, most likely due to different sources of variation in pronoun use for the two groups. The case series approach showed evidence for the distinction of input and output phonological WM buffers. Implications from the two approaches will be discussed.