Two recent computational models of reading development propose that irregular words are read using a combination of decoding and lexical knowledge but differ in assumptions about how these sources of information interact, and about the relative importance of different aspects of lexical knowledge. We report developmental data that help to adjudicate these differences. Study One adopted a correlational approach to investigate the item-level relations between the ability to read a word aloud, general decoding ability, and knowledge of the word’s phonological form (lexical phonology) or meaning (lexical semantics). We found that the latter three factors all influenced accuracy of oral reading. We observed trends indicating that the impact of differences in decoding skill and lexical knowledge were more prominent for irregular words. Study Two comprised two experiments in which novel irregular words were taught; in Experiment 1 we compared phonological to no pre-training, while in Experiment 2 we compared phonological to phonological plus semantics pre-training. Exposure to the phonological form of the word had a substantial impact in the early stages of learning, while the impact of adding semantics was more modest and emerged later. Our findings provide strong evidence that irregular words are read using a combination of decoding and lexical knowledge, with a greater contribution from lexical phonology than lexical semantics. Computational models of learning to read are currently unable to fully account for our data, therefore we propose some modifications. We advocate an instructional approach whereby phonics and vocabulary teaching are combined to support irregular word reading.