AI noise reduction vs spectral gating

Every audio tool has a button that says noise reduction. Behind that button live two completely different technologies, and they fail in completely different ways. Knowing which one you are using explains most denoising disappointments.

Spectral gating: the classic

Spectral gating learns what the noise looks like (from a quiet passage, or estimated statistically), splits the audio into frequency bands, and turns down any band that drops near that noise profile. It has powered audio restoration since the 1990s and runs in milliseconds.

Where it wins: steady, predictable noise. Tape hiss, mains hum, a constant fan. Mild strengths on music, where you want texture kept.

Where it fails: push it hard and you get the famous underwater warble, musical noise, chirpy artifacts in the gaps between words. It cannot tell voice from noise; it only knows loud bands from quiet ones. Noise that changes (traffic, typing, a cafe) defeats the profile it learned.

Neural models: trained separation

Models like DeepFilterNet were trained on thousands of hours of paired clean and noisy speech. They do not gate frequencies; they estimate what the speech itself looks like and reconstruct it, which means they keep working when the noise is loud, messy, and changing, at levels where spectral gating has already collapsed into warble.

Where they fail: they are speech specialists. Feed one a music track and it will politely try to delete the instruments, since to a speech model a guitar is noise. And at extreme noise levels they start reconstructing rather than recovering, which can cost quiet syllables (we wrote up that limit separately in why AI cannot always recover speech).

Which one you should use

This is exactly how our auto mode decides: content analysis picks the method (speech leans AI, music leans spectral at reduced strength), and your measured noise floor sets how hard it works. Both engines are also directly selectable in the tuning panel, because analysis is good and ears are better.

One practical tip that outranks the choice

Whichever method: moderate strength twice beats maximum strength once. Maximum settings are where both technologies produce their signature artifacts, and artifacts are harder to listen to than mild residual noise. Our heavy-noise demo used a second pass at 85% rather than one pass at 100%, for exactly this reason.