The estimate of vocabulary size here is based on the number of unique words used. This seems like it is strongly biased: if two artists have the same size vocabulary, but one has released more albums and thus used more words, that artist will probably have used more unique words. To underscore this point, the number of unique words used by Aesop Rock is half of the estimated vocabulary size of the average college student, although to be fair that estimate is the number of words that an individual can recognize, not the number of words they use. (Edit: the bias is somewhat mitigated by the fact that the same number of words is used to estimate the vocabulary for each artist, but the bias is not dependent on sample size alone but also upon the size of the artist's underlying vocabulary; see my comments below.)
The underlying problem is one of estimating the cardinality of a multinomial distribution given a fixed number of samples. In isolation this problem is ill-posed, since it is always possible that there is a word in a given lyricist's vocabulary that he uses with very low frequency and that is unlikely to appear in any sample, but with appropriate prior information it may be possible to obtain an accurate estimate.
This is not my field, but a brief Google Scholar search shows that there are several papers on estimating vocabulary size, or equivalently, estimating the number of species based on sampling. There is a somewhat dated review (http://cvcl.mit.edu/SUNSeminar/BungeFitzpatrick_1993.pdf) that details some methods of estimation (in this case, I believe we are in the domain of "infinite population, multinomial sample" with unequal class sizes). The paper notes that there is no unbiased estimator available without assumptions on the distribution of word use frequencies, but some of the proposed estimators may be more accurate than the naive estimate used here.