steinbro / hyperviseme Goto Github PK

A "viseme" is a visual phoneme, the visual rather than aural information conveyed when speaking -- in simple terms, lip reading. While there are about forty phonemes in the English language, there are only about a dozen visemes distinguishable by the average viewer-listener, making "homovisemes" (words with the same lip-reading appearance) far more common than homophones. Animators have known this for decades, only needing to illustrate a handful of mouth shapes to make dialogue look plausible on cartoon characters.

Exploiting the quantity of homovisemes for humor has gained popularity recently in creations like Bad Lip Reading.

The Carnegie Mellon Pronouncing Dictionary provides a word-to-phoneme mapping. Combining this with a phoneme-to-viseme mapping from Lee and Yook ("Audio-to- Visual Conversion Using Hidden Markov Models," 2002), we can generate a word- to-viseme mapping. We then invert this mapping to get a homoviseme list.

Give it a try:

make hyperviseme phrase="crazy"

Some nice examples:

• yale college -> hancock itch
• the look of love -> tycoon of knife
• one more time -> wine boredom
• time to check out -> subdue jackass

BSD licensed.

Daniel W. Steinbrook [email protected]