It's weird, but I actually had a very similar idea a few years back! I struggled implementing it and left it half done. I was trying to work directly with the letter frequencies; using a Huffman tree is much nicer. Instead, I tweaked it to generate random Sherlock Holmes paragraphs. It gave the best results with Markov state strings of size equal to the average word size.

It's great to see that it gives a really nice output. Also, your paper is a breeze to read.

I ended up making this[0], which is a Diceware-like with more words, and I just added substitutions to see what it would look like.

At entropy 75, here are five from Molis Hai:

look had can's asing his, nativa "Hemmer,'" Charactiour roopiaties more folks and conful never succepth myriage; by! Let taughter, any mind shu goost out uponsibing of?" "Your," rem

And five passphrases with 3 substitutions:

ov&rload grXas<proof investing Men,ts peartrees Yamages automata adVredGstudy/ng chloroformed exh>linYFassessable gGla rumbOs draughtBmanship

It is hard to determine which is easiest to remember, however. We should perform tests; maybe ask a large number of users to remember one passphrase, and send them a mail three days later asking them to input it. If that sounds like fun to you, I can help.

[0]: https://github.com/espadrine/passphrase