If you look at the latest research in neural network based approaches to text-to-speech, you'll find plenty of examples of realistic synthetic voices. However, due to the nature of the available training data (audiobooks), these models 1) are restricted to reading text in a narrative style and 2) aren't generally trained on speech of recognizable public figures. There was a paper a while back that was trained on TED talks and produced fairly convincing voices, some of which you may recognize (Bill gates, Stephen Wolfram and some others - see here https://sjvasquez.github.io/blog/melnet/) Definitely not 100% convincing, but I wouldn't call it absolute trash.