The last Beatles song, finished with machine-learning separation

What it does

See machine learning used as restoration, not generation: one narrow problem solved, then human performance and production finish the song.

Who it helps

People curious about AI in music — and tired of the “AI Beatles song” label.

How it was made

Models

MAL source separation · WingNut Films — documented

Tools

Source separation · Recovering Lennon’s vocal from cassette and piano

Studio production · New performances and finishing

1. A 1995 attempt failed: the piano and cassette noise masked Lennon’s voice, and the technology of the day could not separate them cleanly.
2. Peter Jackson’s WingNut Films team used MAL machine-learning separation to isolate the real vocal from the demo.
3. McCartney and Starr then completed the recording — new parts, George Harrison’s 1995 guitar, strings and production — around that vocal.

What the person did

The model solved one narrow problem and got out of the way; the song is finished by human performance and production. It is machine-learning restoration, not a synthetic Lennon voice.

What we know

Creator-reported: The official release, the 12-minute making-of short film and the label’s technical credits; MAL separated an authorised archival performance.

Access checked — 2026-09-28 · Opened the official release page and the short film on 2026-09-28; both videos are embed-restricted, so the entry stays link-only.

Can you use it?

The recording stays with its rights holders; we link only and host no audio or frames. The cover is an original Made With Models illustration, not a photo from the session. The song is machine-learning restoration of an authorised archival performance — not a synthesized Lennon voice.

About this project

Picked by Made With Models. Added 2026-09-28. Updated 2026-09-28.

Did you use this? Tell us how it went.