How AI helped the Beatles finish Now and Then

The voice at the centre of the Beatles’ Now and Then came from a cassette John Lennon had recorded at home. Machine learning helped extract that existing performance so the group could build a finished record around it. Co-producer Giles Martin later demonstrated the separation publicly, playing the original material and its isolated parts.[2]

Released on 2 November 2023, the song had stalled because Lennon’s piano and singing were inseparable in the demo. The surviving Beatles could add instruments, but could not obtain the vocal clarity they wanted.[1]

Giving the mixer separate things to control

Emile de la Rey, who led the machine-learning work, described MAL’s source-separation network as creating detailed masks over the audio. These select the target sound while suppressing other components. For Lennon’s demo, the target was his singing, separated from piano and recording noise. This describes a particular engineering task: extracting a usable vocal from sound already captured on tape.[3]

Imagine raising a quiet voice in a recording that also contains a loud instrument. Turning up the whole recording raises both. Separating them gives the mixer independent control, allowing the voice to sit beside newly recorded accompaniment.

The sound team faced a related difficulty on Get Back. Mono recordings could contain conversations, rehearsal music and crew noise together. In their 2022 interview, the editors explained how developing MAL let them rebalance those elements and make dialogue usable within the film. Better separation changed which moments the filmmakers could present intelligibly, connecting an audio tool to the documentary’s storytelling.[4]

The source recording still mattered

The restoration was iterative. De la Rey told Sound & Vision that an early separated vocal still lacked clarity. A later transfer had fuller sound but had already undergone processing and editing. The decisive material was a further transfer of the original cassette without that prior treatment, from which he could isolate a coherent vocal. The quality and history of the input remained part of the solution.[3]

A recovered performance became a new record

The finished track includes Harrison’s guitar from the 1995 sessions, new drums from Starr, and bass, piano and guitar from McCartney. McCartney and Giles Martin produced it; Spike Stent mixed it. A string arrangement by McCartney, Martin and Ben Foster added another newly performed element.[1]

Martin’s 2025 demonstration also identifies who made the creative decisions. McCartney led the return to the song and shaped its final arrangement, including removing a section from the demo. An isolated vocal made that work possible, while the musicians and producers decided what should surround it. The published demonstration is a useful way to hear the distinction between separation and arrangement.[2]

The recording subsequently won Best Rock Performance at the 2025 Grammy Awards. That is recognition of the released performance, rather than a technical score for MAL, but it gives the case an unusually tangible outcome: the recovered material became a record audiences and musicians could judge.[5]

Making more of an archive audible

The same family of tools also contributed to the 2024 restoration of Michael Lindsay-Hogg’s Let It Be film. Park Road Post worked from the original 16mm negative and used MAL in remastering its sound, with the director’s support. The project brought the original documentary back into release alongside the newer Get Back interpretation of the sessions.[6]

These projects show a useful role for AI in cultural preservation: improving access to performances and conversations already recorded. The recovered material gives musicians and filmmakers more to work with, while their decisions determine how it reaches an audience.[4][6]

Report an error or suggest a correction