[vunderba]

Nice job.

A lot of early “fantasy word” generators worked in a similar way - by using n-grams at a character (or character chunk) level rather than the word level even if the breaks didn’t align with syllables.

I don't know if it was ever done - but adapting a Markov model built out against a corpus of word IPAs treating each phoneme as a symbol in the chain would have produced better results as well.