A noisy café. You're trying to make out two quiet conversations: one from the table by the window (black hole mergers), the other a barely audible whisper from the far corner (echoes of the universe's birth). But the hiss of the espresso machine and street noise (detector jitter) drown everything out. New work teaches a computer to separate these sounds.

The authors built a neural network with a special architecture — a multi-level autoencoder with multiple 'heads.' First, it learns to distinguish the gravitational hum from detector noise LIGO-Virgo-KAGRA. Then a statistical method steps in — Markov chains (an algorithm that iterates through possibilities until it finds the most plausible one). It decomposes the cleaned signal into two components: waves from merging black holes and neutron stars, and waves from processes in the young universe.

In a blind test against the standard pygwb pipeline, the new approach more accurately determined both the signal strength and its spectral slope. The key result: it could isolate the cosmological gravitational-wave background (as flat as white noise) at a level of 9.7×10⁻¹⁰ — even when it was hidden under a louder astrophysical hum. It's like hearing a whisper in a café when someone nearby is talking loudly.

So the noisy café is no longer a hindrance. The computer has learned not just to hear both conversations but to understand who said what. Now we wait for real data, and perhaps we'll hear the voice of the youngest universe for the first time.