Tacotron 2 but using CBHG-based for Postnet?

Hello, I want to ask something.

Why this repo using CBHG-based for postnet instead using 5-layer CNN in the tacotron model? Is there any benefit of using it? In my opinion, using 5-layer CNN should be more simple rather than CBHG that are more complex in computation. I will try it later but before that, I want to ask first. Is someone has ever try using 5-layer CNN in this repo?