I can't find the reference anymore, but there is a body of research which takes human motion and extracts a language from it including grammar, dialog acts, words, phonemes.

The possibilities of the video/body movement equivalent of speech recognition would open up a massive, incredibly interesting new dataset to wrangle.