New AI Cleans Up Scratchy Old Orchestral Recordings

Old recordings of orchestras — the tinny, crackling symphonies from the 1900s to the 1930s — have never been properly cleaned up by AI. That's because restoration AI normally learns by comparing a damaged file to its clean original. For century-old records, no clean original exists. You can't teach a machine what a perfect version sounds like if nobody ever recorded one.

A team of six researchers solved that by working backwards. They took modern, clean recordings and deliberately made them sound old — adding the hiss, scratches, muffled highs and narrow range of early recording equipment. That gave the AI thousands of before-and-after pairs to learn from. The AI then works in what's called 'latent space,' a compressed shorthand version of sound, roughly like editing a photo's thumbnail instead of every pixel. It only handles full orchestras, not just solo instruments, which earlier attempts mostly stuck to.

Test results were strong. Their model beat existing restoration tools on both computer scoring and real human listeners, who judged the cleaned-up audio as closer to how a live orchestra should sound. The team also released a 9.3-hour free library of historical classical music for testing, alongside their code and sample audio clips — meaning other researchers, and eventually music labels and streaming services, can build on it.

The honest catch: the AI learned from simulated aging, not the actual quirks of a 1925 gramophone. Real vintage discs may still carry artifacts it never saw. It's also a research paper submitted to a conference, not a product you can download today. Still, the path from 'never properly attempted' to 'measurably better than everything else' is a big jump — and the free dataset is what makes the next round of progress likely.