When Snowd4y, a parody rapper from Toronto, released his track “Wah Gwan Delilah” featuring Drake via Soundcloud on Monday (June 3), it instantly went viral.
“This must be AI,” one commenter wrote of the song. It was a sentiment shared by many others, especially given the tracks ridiculous lyrics and the incredible sound quality of Drake's vocals.
To date, the two rappers have not confirmed or denied the AI rumor. Although Drake posted the track on his Instagram story, it's no confirmation that the vocals in question aren't AI-powered. (As we learned during Drake's recent beef with Kendrick Lamar, the rapper isn't afraid of fake voices.)
To try to get to the bottom of the “Wah Gwan Delilah” mystery, Advertising sign contacted two companies specializing in AI audio detection to review the piece. The answer, unfortunately, was not very satisfactory.
“Our first analysis reveals SOMETHING traces of it [generative] Artificial Intelligence, but there seems to be a lot of combination involved,” he wrote Romain SimiandChief Product Officer of Ircam Amplify, a French company that creates audio tools for rights holders, in an email response.
Larry Mills, senior sales representative at Pex, which specializes in tracking and monetizing music usage on the web, also found mixed results. He said Advertising sign the Pex research and development team “ran the song [their] VoiceID match” and that “Drake's voice in the verse 'Wah Gwan Delilah' doesn't match Drake's voice that well…[as his voice on] official publications [does]but it's close enough to confirm that it could be Drake's own voice or a good AI copy.” Notably, Pex's VoiceID tool alone isn't enough to definitively distinguish real voices from AI , but detecting differences between the singer/rapper's voice on “Wah Gwan Delilah” and his other, officially released songs could indicate some level of AI manipulation.
How to check for AI in songs
There are many types of tools currently in use to distinguish between AI-generated music and human music, but these nascent products are still developing and not definitive. Like Pex's Yakub Galka recently wrote in a company blog post on the topic, “Identifying AI-Generated Music [is] a particularly difficult task”.
Some detectors, like Ircam's, recognize music with artificial intelligence using “artifact detection,” meaning they spot parts of a piece that are off-base from reality. A clear example of this is seen with AI generated images. Early AI images often featured hands with extra or malformed fingers, and there are some detection tools to spot these inaccuracies.
Other detectors rely on reading watermarks embedded in AI-generated music. Although these watermarks are not perceptible to the human ear, they can be detected by some tools. Galka writes that “since watermarking is intended to be detectable by watermark detection algorithms, such algorithms can also be used to show how to remove or modify the watermark embedded in the audio so that it is no longer detectable” — which considers as a major flaw with this detection system.
Pex's method of using VoiceID, which can determine if a singer matches between multiple recordings, can also be useful in AI detection, though it's not a clear answer. This technology is especially useful when users go online and release random tracks featuring Drake's vocals, whether they're leaked songs or fake AI. With VoiceID, Pex can tell a rights holder that their voice was detected on another track that may not be an official release by them.
When VoiceID is paired with the company's other product, Automatic Content Recognition (ACR), it can sometimes determine whether a song uses AI vocals or not, but the company says there isn't enough information for “Wah Gwan Delilah” to complete a full ACR check.
Parody's role in AI music
While it cannot be determined beyond doubt whether “Wah Gwan Delilah” contains AI vocals, parody songs in general have played an important role in popularizing and normalizing AI music. This is especially evident on TikTok, which is full of so-called “AI Covers”, which pair famous singers with amazing songs. Popular examples of this trend are Kanye West singing Natasha Bedingfield's 'Pocket Full of Sunshine', Juice WRLD Coldplay's 'Viva La Vida', Michael Jackson singing George Michael's 'Careless Whisper' and more .
Most recently, AI comedy took center stage with Metro Boomin's SoundCloud-released track “BBL Drizzy” — which sampled an AI-generated song of the same name. The track poked fun at Drake and his alleged “Brazilian Butt Lift” during the rapper's beef with Lamar, and in the process, became the first major use of an AI-generated sample. Later, Drake and Sexyy Red sampled the original AI-generated “BBL Drizzy” on their own song, “U My Everything,” taking “BBL Drizzy” to new heights.
from our partners at https://www.billboard.com/music/rb-hip-hop/drake-wah-gwan-delilah-vocals-real-or-ai-deepfake-1235703089/
