How does it generate voice when it's based on a text based model? I read the readme but afraid I didn't understand