Languages have been shown to optimize their lexicons over time with respect to amount of phonological signal allocated to words: words that are on average less predictable tend to have more segments. But not all segments are equally informative: listeners identify words from the speech stream incrementally, continually updating their lexical search as the phonetic signal unfolds. As a consequence, segments earlier in words contribute on average more disambiguating information to lexical access than later segments, and contribute more or less depending on what similar lexical competitors exist. We expect then that languages should not only optimize the total number of segments allocated to different words, but also how informative those segments are in disambiguating from existing competitors in the lexicon. Here I’ll show data from a range of languages that this is the case: words that are on average less predictable have relatively more informative early segments, while tending to preserve a ‘long tail’ of more redundant later segments.
Second, I’ll review our recent work suggesting that this asymmetry in segment information distribution acros the word may influence the evolution of phonological rules which impact lexical identification. In a typologically-balanced sample of 50 languages, we find that phonological rules which neutralize lexical distinctions (e.g., word-final obstruent devoicing in German) are common at word-ends, but very rare at word-beginnings.