Functional load (FL) is an information-theoretic measure that captures a phoneme’s contribution to successful word identification. Experimental findings have shown that it can help explain patterns in perceptual accuracy. Here, we ask whether the relationship between FL and perception has larger consequences for the structure of a language’s lexicon. Since reducing FL minimizes the risk of misidentifying a word in the case where a listener inaccurately perceives the initial phoneme, we predicted that in spoken language, where perceptual accuracy is important for successful communication, the lexicon will be structured to reduce FL in auditorily confusable initial phonemes more than in written language. To test this prediction, we compared FL of all initial phonemes in spoken and academic written genres of the COCA corpus. We found that FL in phoneme pairs in the spoken corpus is overall higher and more variable than in the academic corpus, a natural consequence of the smaller lexical inventory characteristic of spoken language. In auditorily confusable pairs, however, this difference is relatively reduced, such that spoken FL decreases relative to academic FL. We argue that this reflects a pressure in spoken language to use words for which inaccurate perception does minimal damage to word identification.