Noise reduction and audio restoration basics: what denoisers fix and what they damage

Every home recording has some noise: fan hum, preamp hiss, a fridge, vinyl crackle. Restoration tools can remove a lot of it, but every pass takes something from the music as well. This guide explains how the two families of denoiser work, what each one damages, and the order we'd work in. Sources were read on 30 Oct 2026.

1. Name the noise first

Different noise needs different tools, so identify it before you reach for a plugin.

• Broadband hiss (preamp, tape, room tone): steady, spread across the spectrum. This is what a noise-profile denoiser is built for.
• Hum and buzz (the mains frequency and its harmonics): tonal, so narrow notch filters or a dedicated de-hum usually beat a broadband denoiser.
• Clicks, pops and crackle (vinyl, digital glitches, mouth noise): short events, handled by a de-click tool, not a denoiser.
• Variable noise (traffic, voices, a room that changes): the hard case, and where AI tools are most tempting.

2. Classic denoisers: learn the noise, subtract it

The classic method is spectral: you show the tool a stretch of noise alone, it learns that noise's profile, then it turns down whatever matches across the track. Audacity's Noise Reduction is a clear free example. It needs a noise-only selection of at least 2048 samples (about 0.05 seconds at 44.1 kHz), and longer is better. Its controls are reduction amount (default 6 dB), sensitivity (how much counts as noise) and frequency smoothing, and the manual advises using the lowest reduction that gives an acceptable result.

The damage has a name. Push too hard, or use a profile that doesn't match the whole track, and you get random bursts of short tones the manual calls "musical noise", "bird song" or "tinkly-bells", which it says "can be even more objectionable than the original noise". It also says results are poorest when the noise is loud or variable, or when the music isn't much louder than the noise.

The single most useful feature is the Residue option, which plays only what the tool is removing. If you can hear your guitar's pick attack or a singer's breath in the residue, you are cutting music, not noise.

3. Clicks are a separate job

Audacity's Click Removal shows the same trade-off in miniature. Set the threshold too low and it falsely detects clicks and damages the audio; too high and clicks stay. It also has limits: clicks of 10 ms or longer may not be removed, and soft rapid crackle is better treated with Noise Reduction. Percussive material is the risk here, because a sharp hi-hat or a plucked string looks a lot like a click to a detector.

4. AI denoisers: predict the clean signal

AI denoisers learn from training data what clean audio, usually speech, tends to look like. RNNoise, an open-source library under the BSD-3-Clause licence, is a good one to understand because its method is published. Its paper describes a hybrid: a neural network estimates gains for each frequency band, and a traditional pitch filter removes noise between the harmonics of the voice. The author reports clearly higher quality than a traditional spectral estimator, running in real time at 48 kHz.

Commercial tools go further. Adobe describes Enhance Speech as a "free AI filter for cleaning up spoken audio". Note the word spoken.

What we think: that is the key limit of AI denoisers for musicians. Most are trained and tuned for voice, so they treat whatever doesn't sound like speech as something to remove. On a podcast that's the point. On a song, sustained pads, reverb tails, cymbals, breathy singing and guitar string noise are exactly the "non-speech" detail a speech model may thin out or smear. Where a classic denoiser leaves a bit of hiss or a few chirps, an AI tool can produce something worse for music: a clean-sounding but slightly synthetic voice, consonants that sound reconstructed, and room sound that simply vanishes. Because the output sounds tidy, the damage is easier to miss. That is our listening judgement, not a measured test, and the best tools improve every year, so check them on your own material.

5. A safe order of work

1. Fix the source. Re-record if you can. Turn off the fan, move the mic closer, gain-stage properly. No plugin beats a quieter room.
2. Work on a copy. Keep the untouched file, saved as WAV (see audio file formats).
3. Remove tonal hum first with notch filters, then clicks, then broadband noise. Each tool finds its target more easily once the others are gone.
4. Use small passes. Two gentle passes usually sound better than one heavy one. Start near the defaults.
5. Listen to the residue or a difference track. Music in the residue means back off.
6. Use a mix control. Blending some of the original back in often keeps the life with most of the noise gone. If a tool has no mix knob, do it with two tracks.
7. Use AI tools on speech-like parts, such as a spoken intro, a dry vocal or a field recording of dialogue, and judge them soloed and in the mix. For editing specific words, see AI vocal editing.
8. Compare at matched loudness. A denoised track is often quieter, and quieter often sounds "cleaner" for the wrong reason.

A little noise under a full mix is rarely heard. Audible processing usually is. When in doubt, take less out.