Over the last century, evidence has accumulated suggesting that language structures evolve to maintain a balance between effort and communication accuracy. However, both ‘effort’ and ‘accuracy’ are influenced by the particular processes involved in encoding and decoding the speech signal. When listeners hear the beginning of a word, they begin narrowing down hypotheses about what word they are hearing from the very first segment, rather than waiting until the end of the word. As a consequence, early segments contribute more disambiguating information than later segments. If languages evolve to optimize communication efficiency, we expect that informative segments should be concentrated early in words, because that's where they can do the most work in disambiguation. Here I'll show data showing that words that are less predictable do in fact concentrate highly informative segments early, while also preserving a longer tail of redundant segments. Second, I’ll show evidence that the lower information of late segments influences the evolution of phonological rules. In a sample of 50 languages, we find that phonological rules which neutralize lexical distinctions are common at word-ends, but rare at word-beginnings, where neutralization would more negatively impact lexical identification.