On November 2, 2023, The Beatles released “Now and Then,” billed as the band’s final song, built around a John Lennon vocal recovered from a home demo he recorded in the late 1970s. The track was issued worldwide as a double A-side paired with the band’s 1962 debut single “Love Me Do.”
The obstacle had stood for decades: on Lennon’s cassette, his voice and piano were recorded to the same track over a hiss of background noise, and earlier attempts to separate them in the 1990s had failed. The breakthrough came from machine-learning audio tools built by Peter Jackson’s WingNut Films team. Using a technology they called MAL, originally developed to de-mix the mono soundtrack for the 2021 Get Back documentary, engineer Emile de la Rey applied the same technique to Lennon’s recording, isolating a clean vocal from the piano and tape noise. McCartney, Starr, and archival Harrison parts completed the song.
The release was a high-profile, sympathetic use of AI in music: rather than generating a synthetic voice, the tools extracted and preserved a real performance that would otherwise have been unusable. It stood in contrast to the more contested uses of voice AI emerging at the same time.
Why business readers should care: Now and Then showed AI audio separation as a restoration tool, a constructive counterpoint to the deepfake-voice anxieties dominating the music conversation.