In this paper we investigate whether homophones have shared (e.g., Dell, 1990; Levelt, Roelofs, & Meyer, 1999) or independent (e.g., Caramazza, Costa, Miozzo, & Bi, 2001) phonological representations. We carried out a homophone reading aloud task with low frequency irregular homophones and matched low frequency irregular non-homophonic controls. The 'Shared Representation' view predicted a homophone advantage: homophones should be read faster than their matched controls because the low frequency homophone inherits the frequency of its high frequency partner. The 'Independent Representation' view predicted neither an advantage nor a disadvantage: performance should be governed by the homophone's specific-word frequency. Results showed that low frequency homophones were read aloud slower than non-homophonic controls. Results were confirmed with an independent database of reading latencies (Balota, Cortese, Hutchison, Neely, Nelson, Simpson, & Treiman, 2002). Additionally, attempts to simulate the homophone disadvantage effect using current computational models of reading aloud were all unsuccessful. The homophone disadvantage effect constitutes, therefore, a new challenge for all computational reading models to date.