海角大神

Artificial intelligence has been used to extract John Lennon鈥檚 voice from an old demo to create 鈥渢he last Beatles record,鈥� decades after the band broke up, Paul McCartney said Tuesday.

Mr. McCartney, an octogenarian, told the BBC that the technology was used to separate the Beatles鈥� voices from background sounds during the making of director Peter Jackson鈥檚 2021 documentary series, 鈥淭he Beatles: Get Back.鈥� The 鈥渘ew鈥� song is set to be released later this year, he said.

Mr. Jackson was 鈥渁ble to extricate John鈥檚 voice from a ropey little bit of cassette and a piano,鈥� Mr. McCartney told BBC radio. 鈥淗e could separate them with AI, he鈥檇 tell the machine 鈥楾hat鈥檚 a voice, this is a guitar, lose the guitar鈥�.鈥�

鈥淪o when we came to make what will be the last Beatles record, it was a demo that John had that we worked on,鈥� he added. 鈥淲e were able to take John鈥檚 voice and get it pure through this AI so then we could mix the record as you would do. It gives you some sort of leeway.鈥�

Mr. McCartney didn鈥檛 identify the name of the demo, but the BBC and others said it was likely to be an unfinished 1978 love song by Mr. Lennon called 鈥淣ow and Then.鈥� The demo was included on a cassette labeled 鈥淔or Paul鈥� that Mr. McCartney had received from Mr. Lennon鈥檚 widow, Yoko Ono, the BBC reported.

Mr. McCartney described AI technology as 鈥渒ind of scary but exciting,鈥� adding: 鈥淲e will just have to see where that leads.鈥�

The same technology enabled Mr. McCartney to 鈥渄uet鈥� virtually with Mr. Lennon, who was murdered in 1980, on 鈥淚鈥檝e Got a Feeling鈥� last year at Glastonbury Festival.

Holly Herndon, a multidisciplinary artist with a doctorate in composition from Stanford University, used nascent AI machine technology on her last album, 2019鈥檚 鈥淧roto,鈥� and developed Holly+, an online protocol that allows the public to upload tracks to be reinterpreted and performed by a deepfake version of her voice. She theorizes that the Beatles鈥� recording was likely created using a process called 鈥渟ource separation.鈥�

鈥淪ource separation has become much easier to do with machine learning. This allows you to extract a voice from a recording, isolating it so that you might accompany it with new instrumentation,鈥� she explains.

That differs from a deepfake vocal. 鈥淎 deepfake is an entirely new vocal line spawned from a machine learning model trained on old vocal lines,鈥� she said. 鈥淲hile it does not appear to be happening in this example, it is now possible to spawn infinite new media from analyzing older material, which is a similar process, in spirit, to this song.鈥�

Mr. McCartney is set to open an exhibition later this month at the National Portrait Gallery in London featuring previously unseen photographs that he took during the early days of the Beatles at the start of 鈥淏eatlemania,鈥� when the band rose to worldwide fame.

The exhibition, titled 鈥淓yes of the Storm,鈥� showcases more than 250 photos Mr. McCartney took with his camera between 1963 and 1964 鈥� including portraits of Ringo Starr, George Harrison, and Mr. Lennon, as well as Beatles manager Brian Epstein.

This story was reported by The Associated Press. AP writer Maria Sherman reported from Los Angeles.