I tried to train a Piper voice based on the game assets from Portal2. It turned out okay-ish, though my GTX1660Super really limits the batch size and training speed. I sadly don't have more compute power available nor the time to leave it running in the background.
Here are some examples: [https://go.wetransfer.com/t-nhqOLcDQkh]
Her voice makes hiding imperfections pretty easy. Since we can't acquire more recordings, the only option left, is to change the training process.
If someone with a better GPU want's to give this a shot, I can send over the dataset, which I have already prepared. Since I don't know about the legality of sharing these fiels (dataset and onnx) I will not be posting them publicly, as I don't own the copyright (though they are available on the internet).