Did you read the paper? They intentionally steered the quality to ensure they sound fake. Their generated speech is "very easy to detect" according to the reference at the end of the paper.