I like this idea, but I wonder if the potential for confusion or even offence is too high for it to be workable.  I'm thinking of https://en.wikipedia.org/wiki/Sudanese_teddy_bear_blasphemy_case.

> we can swap these with colors or other words to make it little bit less confusing (e.g. "blue-star"). 

See https://pypi.python.org/pypi/petname and https://www.npmjs.com/package/human-readable-ids, which both use animals.

As an alternative, I wrote a program that takes the distribution of trigrams from an English dictionary, and statistically generates a Markov chain of such overlapping trigrams that look almost entirely unlike English words but often are strangely pronounceable, for instance:

umirckbysag mpspiarefor doptinenchc lymdeotmicn gclyowdhoki
gavotlilmod hamdipicpto sriagflyori nstsdumotdo biclachiesf
dsalcleccod  lreamyryazi tbrtnmchnbl gsciffsucec lveadjeortd
itiasglucpa   nctyrsifesk   asbeazeimst rvalvallwsh gbrcarpirgo
ootsjamakid ybawffobrcr wsbiopradsn pouidbrmcif nytblplabio
bytramojetw hdenleraloa crymolduanu nutoloymctr ofanjewsstm

We've been using them for labels and such, and of course in principle this method should occasionally generate actual existing words, including NSFW words in English and possibly in other languages, but it turns out that the probability is very small, for what it's worth.