New AI Cleans Up Old Records and Tapes — With Sliders You Control

Most audio cleanup software works like a one-button filter: you feed in a damaged recording, and it hands back a single "fixed" version. The problem is that the machine quietly decides what counts as noise and what counts as music. Feed it a scratchy blues record, and it might scrub away the very warmth that made the recording special. A new tool called RESTORE, from researchers Meiying Chen, Benjamin R. Thompson and Michael C. Heilemann, takes a different approach.

RESTORE breaks a messy recording into six separate layers: vocals, instruments, broadband hiss, sudden pops and clicks, and an "everything else" leftover. Think of it like a mixing board with individual volume sliders — you can turn down the hiss without touching the singing, or ease off the clicks while keeping the vinyl crackle you actually like. It also fills in missing high frequencies, which is what makes old tapes sound dull and far away. The researchers say it noticeably improves quality on historical recordings compared to existing methods, and that users' preferences lined up strongly with the controls they were given.

The clever part is honesty. When AI adds or invents sound — filling in frequencies that were never recorded — RESTORE keeps that invented layer separate and visible, so you can hear exactly what was generated and turn it off if you don't trust it. That matters as AI-generated music and voices become harder to spot.

Speed is the other headline. RESTORE processes audio about 50 times faster than real time on a single graphics chip, so an hour-long record takes barely more than a minute. That puts serious restoration within reach of hobbyists, archivists, podcasters and family historians digitizing old cassettes — not just big studios with expensive software.