This paper will examine the production of ambiguous words (homophones e.g., pair, pear) in aphasia. In the psycholinguistic research literature there is debate regarding the representation of homophones. Some authors (e.g., Levelt, Roelofs & Meyer, 1999) assume homophones share one phonological form, in contrast others suggest that each member of the homophone has a separate phonological representation (e.g., Caramazza, Costa, Miozzo & Bi, 2001). This study presents evidence from the treatment of aphasia that addresses this issue. Three single cases of three aphasics with severe anomia, one German and two English, will be presented. The treatment comprised intensive training of picture naming using phonological cues. It was investigated if this pure phonological training could improve naming performance; what pattern of generalisation was observed across pairs of stimuli that were homophonic, semantically related or phonologically related and the duration of maintenance of the improvement. The results of the German single case showed significant generalisation to untreated homophones, but no generalisation to untreated semantically or phonologically related stimuli. The English aphasics showed the same pattern of item-specific effects for the treated homophones, and generalisation to the untreated homophones. It is argued that this supports a single phonological representation but two lemma representations for both members of a homophone pair. The results are interpreted within a discrete model (such as Levelt et al., 1999) and an interactive model of Dell (e.g., 1990).