AFAIK, the state of the art in voice synthesis still involves concatenating recorded snippets. Even fancy neural networks need recordings to train on, and will most likely sound similar to the original voice actor.