Word learning is a central part of language development—all aspects of language intersect at the level of the word. And it is a surprisingly difficult task. In addition to determining the set of referents to which a word should be mapped, learners must also determine the set of acoustic tokens that constitute the same word. This is a nontrivial problem because no two instances of a word are acoustically identical. Both within and across speakers, words may be realized differently due to lexically irrelevant changes in rate, emotion, and voice quality. When words are mispronounced or produced by speakers with different accents, changes can also occur along lexically relevant dimensions. In the case of mispronunciations, the changes are typically unsystematic and should be ignored. In the case of accents, the changes are systematic and should be learned, not only to facilitate future processing, but also because they transmit social information about speakers. Therefore, not all variability should be treated in the same way. To complicate matters more, what counts as lexically relevant can differ across languages. Some acoustic properties that English speakers consider to be irrelevant for lexical identity, like pitch, are lexically relevant in other languages (e.g., tone languages).