[antpls]

That's great, but it still performs 2 times worst than the best supervised model.

Also : "The discriminator itself is also a neural network. We train it by feeding it the output of the generator as well as showing it real text from various sources that were phonemized."

Is the "real text from various sources that were phonemized" a manually labelized database? If Yes, that step is supervised, which makes the whole thing actually supervised to some extent

[exporectomy]

If phonemization is converting words to phonetic symbols, it could be taken from existing human-written dictionaries automatically.