This study addressed the question of exactly which aspects of spelling-sound consistency influence accuracy of reading aloud in surface dyslexic patients with semantic dementia. Oral reading data were obtained from twelve patients on three sets of words that varied in regularity (defined according to grapheme-phoneme correspondences) and consistency (defined according to the pronunciation of word body neighbours). The patients were less accurate for irregular/inconsistent words, which they commonly pronounced in line with sound-spelling regularities, as expected in surface dyslexia. They produced plausible but incorrect responses for some regular as well as many irregular words, suggesting that their reading performance was influenced by sound-spelling relationships not captured by grapheme-phoneme correspondences. On a set of items that varied consistency and regularity independently, the patients showed a large effect of regularity and a smaller but significant effect of consistency in reading aloud. In addition, there was a correlation between degree of semantic impairment and level of reading accuracy for inconsistent items. These findings are discussed in terms of two influential models of reading: the dual-route-cascaded model (Coltheart et al., 2001) and the triangle model (Plaut et al., 1996). It is argued that the triangle model provides a more straightforward account of the relationship between word comprehension and consistency effects in reading. Copyright © Taylor & Francis Inc.