Yes, the amount of work put into this project is ingenious. From all I know, I've tried to build one github repository that used tenserflow and even a nice user interface. But the results were very bad, although the dev argued the program would clone the voice from 5 seconds of raw data.
The neural net algoritm is primitive and does not rely on a huge amount of data. It's training the ai that does the trick. from the get-go I assumed that once all the training is done there would be an option to share an executable so people would just run it on their hardware and expect something nice in return.
But again, Even a couple of words take minutes to be generated, let alone sentences. there is no doubt that 15.ai heavily relies on a great power if it were capable of firing generated lines with a pace of a shotgun.