Columbia's New AI Can Tell a Vinyl Pop From a Scratch

Every time you play an old record, the software cleaning it up faces a choice: smooth over the noise, patch it, or leave it alone. A click, a short burst of static, and steady crackle all need different fixes. Most tools treat every blip the same way, which is why badly restored albums can sound dull and flat — the good music gets sanded down along with the damage.

The Columbia team split the job in two. First, several small specialist models each hunt for one kind of damage — scattered clicks, quick bursts, or thick crackle. Then a second piece acts like a traffic controller, watching that evidence over time and deciding which repair mode the record is in right now. Instead of making one giant all-at-once guess, it narrows the options step by step.

They tested it on 597 fifteen-second clips pulled from 21 recordings. The system scored 0.730 on a 0-to-1 scale for sorting damage into five categories, beating a flatter approach by 0.043 — a small but statistically solid gain. Clean sections and heavily crackled sections were handled reliably. The messy middle ground, like light scattered clicks mixed with other noise, still confused it most.

This is a lab paper, not a product you can buy. The tests used computer-generated damage rather than real worn records, so nobody knows yet how it holds up against a scratched-up thrift-store find. But the direction matters: restoration software that removes noise without removing the music is valuable to archives, reissue labels, and anyone with a crate of old LPs.