Look at *differences* between words rather than at the words themselves?

[Koen G]

ba-cez ce caz-dyz-faz-fe ge-hiz gix-gew-giv gew-few-gaw-few-giv-gow fov-fit-fos-dos-dur fyr-gyr-fas-fer-fis fit ges-has-har-haq-jaq-hap hap

Okay, so first we have to wonder whether the VM would separate each syllable or not. Let's take the first sentence of the VM:
fachys ykal ar ataiin shol shory cthres y kor sholdy

In this transcription (ZL_2), letter counts are:
6 4 2 6 4 5 6 1 3 6
So everything from 1 to 6. This means that either each source text letter has a 1-6 letter output, or not all output is separated by spaces.
Without separation, your output would be:

bacez ce cazdyzfazfe gehiz gixgewgiv gewfewgawfewgivgow fovfitfosdosdur fyrgyrfasferfis fit geshasharhaqjaqhap hap

That's TTR 100%, which would continue in larger sequences because of the progressive and variable nature of the encoding. I also suspect that even doing some kind of reset at every line would result in a similar variation.

I also suspect that even with practice, it would be a pain to read. But that's the case with every advanced code.

I do understand why you're thinking along these lines though, it's definitely an interesting exercise.