Almost all of the theoretical and empirical work on reading aloud has considered only the reading of monosyllables, and so the special problems which arise when one is attempting to give an account of how polysyllabic words and nonwords are read aloud have been thoroughly neglected. Here we begin to remedy this neglect with an exploratory study of this issue from the viewpoint of the dual-route theory of reading. We propose an explicit set of nonlexical rules for the orthographic-phonological translation of disyllabic letter strings which includes procedures for assigning stress and reducing vowels. We show that this set of rules predicts well how people assign stress to disyllabic nonwords and that the naming latencies for English disyllabic strings whose stress violates that predicted by these rules are longer than the latencies for words which obey these rules, especially when the words are low in frequency. We conclude with a consideration of how a particular dual-route computational model of reading, the DRC model, might be extended so as to account for these findings.