Phonological decoding is central to learning to read, and deficits in its acquisition have been linked to reading disorders such as dyslexia. Understanding how this skill is acquired is therefore important for characterising reading difficulties. Decoding can be taught explicitly, or implicitly learned during instruction on whole word spellings and pronunciation. This study describes the design and testing of a new grapheme-phoneme correspondence (GPC) rule learning model (GPC-LM), based on earlier work by Coltheart, Curtis, Atkins, and Haller. It simulates the implicit deduction of GPCs. These learned GPCs are tested in conjunction with the dual-route cascaded model of reading aloud. The new model learns many productive GPCs and achieves good word reading performance without lexical route participation. Nonword pronunciation using the learned GPCs also more closely matched human data than achieved by the connectionist dual-process models (CDP+/++). Despite this, challenges regarding psychological plausibility remain, and are discussed.