It is likely that generalization of implicitly learned sound patterns to novel words and sounds is structured by a similarity metric, but how may this metric best be captured? We report on an experiment where participants were exposed to an artificial phonology, and frequency ratings were used to probe implicit abstraction of onset statistics. Non-words bearing an onset that was presented during initial exposure were subsequently rated most frequent, indicating that participants generalized onset statistics to new non-words. Participants also rated non-words with untrained onsets as somewhat frequent, indicating generalization to onsets that had not been used during the exposure phase. While generalization could be accounted for in terms of featural distance, it was insensitive to natural class structure. Generalization to untrained sounds was predicted better by models requiring prior linguistic knowledge (either traditional distinctive features or articulatory phonetic information) than by a model based on a linguistically naïve measure of acoustic similarity.