Gravitational telescopes listen to the cosmic 'music' — gravitational waves from distant cataclysms. But their recordings are full of 'scratches' — short terrestrial glitches. Previously, assessment was done over the entire track at once, and faint clicks were lost. The new method divides the recording into microscopic fragments and compares each to a library of known noises.

A neural network with perfect hearing was built into the system: it wasn't taught to recognize specific gravitational signals, but it's trained to distinguish silence from suspicious sounds. This 'critic' detects time-stretched glitches with 96% accuracy. For ultra-short glitches, the method is still somewhat rough, but it constructs a heat map of risks, highlighting areas on the spectrogram (a musical score of frequencies) that resemble defects. This way, scientists won't mistake a real gravitational hum for a scratch on the record.