A new algorithm is presented for vocabulary analysis (word detection) in texts of human origin. It performs at 60%-70% overall accuracy and greater than 80% accuracy for longer words, and approximately 85% sensitivity on Alice in Wonderland, a considerable improvement on previous methods. When applied to protein sequences, it detects short sequences analogous to words in human texts, i.e. intolerant to changes in spelling (mutation), and relatively context-independent in their meaning (function). Some of these are homonyms of up to 7 amino acids, which can assume different structures in different proteins. Others are ultra-conserved stretches of up to 18 amino acids within proteins of less than 40% overall identity, reflecting extreme constraint or convergent evolution. Different species are found to have qualitatively different major peptide vocabularies, e.g. some are dominated by large gene families, while others are rich in simple repeats or dominated by internally repetitive proteins. This suggests the possibility of a peptide vocabulary signature, analogous to genome signatures in DNA. Homonyms may be useful in detecting convergent evolution and positive selection in protein evolution. Ultra-conserved words may be useful in identifying structures intolerant to substitution over long periods of evolutionary time.