This paper compares two theoretical positions regarding the mental representation of homophones: first, that homophones have one phonological word form but two grammatical representations (lemmas, e.g., Levelt et al., 1999; Dell, 1990), or second, that they have two separate phonological word forms (e.g., Caramazza et al., 2001). The adequacy of these two theoretical accounts for explaining the pattern of generalisation obtained in the treatment of homophone naming in aphasia is investigated. Two single cases are presented, where phonological treatment techniques are used to improve word retrieval. Treatment comprised picture naming of one member of a homophone pair using a phonological cueing hierarchy. A significant improvement in word retrieval was found for both the treated and the untreated homophones, while there was no improvement for phonologically and semantically related controls. It is argued that the data support a shared representation for homophones at the word form level. However, current theories cannot explain the pattern of generalisation found without the addition of a mechanism for repetition priming (e.g., suggested by Wheeldon and Monsell, 1992) and feedback between word form and lemmas to explain the results.