レポート 2983

When Marc set out to clone the voice of Ariana Grande in January, he didn’t know it would infuriate hordes of his fellow Grande fans.

Marc, who asked to remain anonymous to shield him from further online attacks, wasn’t looking to incite a stan riot. He just wanted to experiment with DiffSVC, an open-source software developed by researchers at The Chinese University of Hong Kong that can replicate real human voices. By training the model with source material from Grande’s song list, Marc was able to make cover songs in her same style and tone.

Marc started out by collecting data of the singer from the internet, grabbing recordings of her voice from a capella tracks, some directly uploaded to YouTube by Grande and others placed there by her fans. He also pulled from high-quality Dolby Atmos tracks, which are designed to make sound seem more three-dimensional. He ended up with 33 minutes of audio, which he and a friend chopped up into 14-second audio files and threw into DiffSVC.