Early reading abilities are widely considered to derive in part from statistical learning of regularities between letters and sounds. Although there is substantial evidence from laboratory work to support this, how it occurs in the classroom setting has not been extensively explored; there are few investigations of how statistics among letters and sounds influence how children actually learn to read or what principles of statistical learning may improve learning. We examined 2 conflicting principles that may apply to learning grapheme-phoneme-correspondence (GPC) regularities for vowels: (a) variability in irrelevant units may help children derive invariant relationships and (b) similarity between words may force children to use a deeper analysis of lexical structure. We trained 224 first-grade students on a small set of GPC regularities for vowels, embedded in words with either high or low consonant similarity, and tested their generalization to novel tasks and words. Variability offered a consistent benefit over similarity for trained and new words in both trained and new tasks.