This paper uses neuropsychological data to differentiate between three models of verbal short-term memory (Baddeley, 1983, 1986; Besner, 1987; Monsell, 1987). The focus is on three tasks: homophone judgments, rhyme judgments and pseudohomophone detection. When lesioned each model predicts characteristic patterns of impairment across these tasks. Thirteen span impaired aphasic subjects were assessed on all three tasks. The patterns of performance unequivocally supported the model proposed by Monsell (1987) which distinguishes input and output phonological buffers. Implications for further research are discussed.