After forming each sentence in your head, you consider it as a stream of syllables, map every syllable to a byte sequence, and OTR-encrypt it before converting it back into syllables and saying that near-gibberish to the person who you are talking to.
Just in case the NSA is listening, you know.
Exchanging the private keys initially should be done by tongue movements when you French kiss someone. This means that everyone has to have French kissed at least once before they can talk to each other.