Emerging phonotactic knowledge facilitates the development of the mental lexicon, as demonstrated by studies showing that infants use the phonotactic patterns of their native language to extract words from continuous speech. The present study provides a computational account of how infants might induce phonotactics from their immediate language environment, which consists of unsegmented speech. Our model, StaGe, implements two learning mechanisms that are available to infant language learners: statistical learning and generalization. StaGe constructs phonotactic generalizations on the basis of statistically learned biphone constraints. In a series of computer simulations, we show that such generalizations improve the segmentation performance of the learner, as compared to models that rely solely on statistical learning. Our study thus provides an explicit proposal for a combined role of statistical learning and generalization in the induction of phonotactics by infants. Furthermore, our simulations demonstrate a previously unexplored potential role for phonotactic generalizations in speech segmentation. (Contains 3 tables and 7 figures.)