Large-corpus natural speech fits a Zipfian distribution

Zipf’s law says that the frequency of words is inversely proportional to rank. The evidence for this is ordinarily drawn on orthographic texts as the linguistic input. Those orthographic studies are robust, but it is long evident that written language is not a one-to-one analog to spoken language. It has not been demonstrated that natural, spoken linguistic content respects a Zipfian distribution. This project leverages recent computational advances and large-scale naturalistic corpora to examine whether spoken language is Zipfian. As a reference, we use the orthographic Brown Corpus (1.012 million words). Here we examine spoken language data from the Buckeye, Santa Barbara, MiCase, and Fisher corpora. These corpora represent 9.003 million words from corpora varying in collection method, formality, and spoken mode. The analytic method employed here uses a recent technique (Clauset et al. 2009 https://doi.org/10.1137/070710111) to examine goodness-of-fit of the distributions and to determine if those distributions are adjudged to be Zipfian. Results affirm the spoken language corpora examined here are Zipfian. Results may be useful in understanding speech and language acquisition, disorders, comparisons with non-English corpora, and machine approaches to language including human–computer interaction.