I'm not very convinced by the article. It introduces a method of obfuscation that could be solved once and for all by a mapping built in like 1h (which is probably also part of the software used for obfuscation itself).
I wouldn't be surprised if some NLP libraries already normalize this kind of text?