It was Nvidia's Tacotron2[0] but now I believe it's NaturalSpeech[1]

[0]https://paperswithcode.com/method/tacotron-2 [1]https://speechresearch.github.io/naturalspeech/