This paper investigates the possibility that phonological knowledge emerges out of learning how to process words. A connectionist model of word recognition and production is presented, and a series of experiments is described in which a network is trained to recognize or produce a small set of words from an artificial or a real natural language. In the process of learning these tasks, the network develops internal, distributed representations of its state at different points during processing. In one set of experiments, the internal representations which emerge during a recognition task are treated as inputs to other networks, where their adequacy as syllable representations is tested. It is shown that the representations (1) support word production as well as recognition, (2) support mutation, insertion, and deletion processes, and (3) are robust to noise in the input. In another experiment, representations which develop during a production task are analyzed using a dimensionality reduction technique. It is shown that two of the dimensions exhibit some of the properties of the tiers of autosegmental analyses.