Noise reduction / spectral repair
Learning what the background noise sounds like and subtracting it, or painting a single cough out of the picture of the sound.
See it
What it is
Two different jobs that live in the same toolbox. Broadband noise reduction learns a profile from a second or two of pure noise (a pause where nobody talks) and subtracts that fingerprint across the whole file: Audacity's Noise Reduction, Adobe Audition's Noise Print, iZotope RX Voice De-noise. Spectral repair treats the spectrogram as an image, so you lasso the chair squeak, phone buzz, or single cough and heal it from the surrounding audio, like the clone stamp in Photoshop. Mains hum gets its own tool that notches 50 or 60 Hz plus every harmonic above it.
Reach for it after you have fixed everything fixable at the source, and before compression, since compression will amplify whatever you leave behind. Always capture the noise profile from the same take and the same room; a profile borrowed from another clip subtracts the wrong thing.
Gotcha: the artifacts are worse than the noise. Push reduction past roughly 10 dB and voices go underwater, gargly, and surrounded by chirping 'musical noise' that no listener has a name for but everyone hears as cheap. Two light passes beat one heavy one, and leaving a little honest hiss is a legitimate choice.
Ask AI for it
Clean up this dialogue recording without artifacts. First locate a genuinely speech-free region in this take and verify it contains no voice, breath, or mouth noise before profiling it; do not assume the head of the file is clean. If no usable noise-only region exists, use an adaptive voice denoiser instead of a learned profile. Apply broadband reduction of no more than 6 to 9 dB. Then detect whether the mains hum here is 50 Hz or 60 Hz, and notch only that fundamental plus the harmonics that are actually audible. Use spectral repair to heal isolated one-off sounds (mouth clicks, chair creaks, a single beep) rather than raising the overall reduction. Preserve the natural tone of the voice: no underwater, chirpy, or gargly artifacts, and leave a low steady noise floor rather than absolute silence between phrases.