First, identify what you actually have
“Noise” is four different problems that need four different fixes. Almost every bad result people get from noise reduction comes from applying one tool to all of them.
Find a few seconds of your recording where nobody is speaking or playing, loop it, and turn it up. What you hear tells you which problem you have — and our free noise-floor check puts a number on it in the browser, alongside whether the file is clipped, which decides the order below:
- A steady hiss or rush, like distant traffic or air — broadband noise. Preamp noise, air conditioning, a computer fan, the room’s own noise floor.
- A steady low tone, often with a buzzy edge — mains hum, at 50 or 60 Hz depending on your country, plus harmonics above it.
- Discrete events — clicks, pops, a brief digital stutter. Different problem entirely; these are damage, not a floor.
- Nothing much, but the take sounds distant or boxy — that is the room, and it is not noise at all. It is your recording plus the sound of where you made it.
You may well have several at once. Treat them one at a time, in the order below.
Broadband noise: hiss, fans, air
Broadband noise is spread across the whole spectrum, at a fairly constant level. The standard approach is spectral noise reduction: the tool learns a profile of the noise from a section where only noise is present, then reduces energy matching that profile across the whole recording.
Capture the profile properly. Give the tool a clean section of noise only — no breath, no chair creak, no distant door. A profile taken from a passage that includes the tail of a word will teach the tool that part of your voice is noise, and it will faithfully remove it.
Then use much less reduction than you want to. Six decibels of reduction is often enough to move noise from distracting to unnoticeable. Twenty decibels will certainly remove the noise, and it will take the top of the voice with it.
A useful trick: set the reduction far too high on purpose, listen to how mangled it sounds, and then back off until it stops. That teaches you where the boundary is on this specific recording much faster than creeping up from zero.
Sometimes the honest answer is an equaliser instead. If the noise is mostly low rumble, a high-pass filter at 80 Hz removes it completely with no artefacts at all — noise reduction is not required, and using it anyway costs quality for nothing. Looking at the average spectrum is usually enough to tell which case you are in: rumble is a slope at the bottom, hiss is a shelf across the top, and hum is a picket fence of spikes.
Hum and buzz
Mains hum sits at 50 Hz or 60 Hz with harmonics stacked above it — 100, 150, 200 and so on, or 120, 180, 240. Because it is a set of precise tones rather than a spread, it can be removed almost surgically with a comb of narrow notch filters, at essentially no cost to the material around it.
Two things matter. First, remove the harmonics, not just the fundamental — the buzzy character people find most irritating lives in the harmonics, and notching only 50 Hz leaves it entirely intact. Second, prefer a tool that tracks the frequency, because mains supply drifts slightly. A static notch at exactly 50 Hz gradually stops covering hum that has wandered to 50.2 Hz, while continuing to remove a sliver of your bass.
If the hum appeared only when you touched something, it is a grounding or cabling problem, and re-recording after fixing it beats any amount of processing.
Clicks, pops and dropouts
These are damage at specific moments: a bad edit, a buffer glitch, a bad cable connection, vinyl surface noise, a mouth click on a close-miked voice.
They are repaired rather than reduced. A declicking tool finds the impulse and reconstructs the missing fraction of a second from the waveform surrounding it — and because the affected region is genuinely tiny, a good repair is completely inaudible.
For a handful of clicks, doing it by hand is often better than automatic detection: zoom in, find the discontinuity, and either redraw the waveform or apply a very short crossfade. Automatic declicking on a whole track will occasionally decide a consonant is a click.
Mouth clicks on a vocal are their own annoyance and respond well to manual repair, which is tedious and worth it on a lead vocal.
Room sound
If your recording sounds boxy, distant, or like it was made in a bathroom, that is the room — the reflections that arrived at the microphone shortly after the direct sound. It is not noise, and noise reduction will not touch it, because it has the same spectral content as the thing you want to keep. It is the thing you want to keep, arriving late.
Dereverb tools exist and have become genuinely good, using models that have learned what speech and music look like with and without a room. They reduce the reflections while leaving the direct sound.
Use them gently, for a reason that is easy to miss: a completely dry voice sounds wrong. Real listeners have never heard a voice with no room at all, and stripping it entirely produces something people find unsettling without being able to say why. Reduce the room until it stops being distracting, then stop.
Of the four problems here, this is the one that processing helps least. It is also the one most cheaply prevented, which is the next section.
The order matters
Do these in this order, because each one changes what the next one sees:
- Repair damage first — declip and declick. A clipped or clicking sample teaches a noise profile the wrong thing about what noise looks like here.
- Remove hum next. It is tonal and surgical, and removing it makes the broadband noise floor easier to profile accurately.
- Then broadband noise reduction, gently.
- Then dereverb, if you need it — on a signal that is now clean, so the model is not trying to interpret hiss as early reflections.
- Then everything else — equalisation, compression, the actual mix.
And do all of it on the individual tracks, before mixing. Once a noisy track is summed with everything else, its noise is entangled with material you want, and the same repair does far more collateral damage. If the session is gone and only the bounce survives, separating it back into parts at least lets you treat the damaged element on its own — inheriting whatever the separation left behind, which is the trade you are making.
Why noise reduction makes things sound underwater
The artefact has a name — musical noise — and understanding it makes it much easier to avoid.
Spectral noise reduction divides the signal into many narrow frequency bands, many times a second, and attenuates each one where it looks like noise. When the reduction is aggressive, isolated bands survive here and there, essentially at random, and each one is a brief pure tone. Hundreds of those per second is the burbling, watery, underwater sound.
It is a symptom of asking for more reduction than the recording supports. The cures are all the same one: reduce less, take a better noise profile, and accept a slightly higher noise floor. A recording with audible hiss sounds like a recording. A recording full of musical noise sounds broken, and listeners forgive the first far more readily than the second.
The part nobody wants to hear
Fifteen minutes before the take beats any amount of repair afterwards. Turn off the air conditioning and anything with a fan. Move the microphone closer, which raises your signal against the room and the noise floor at the same time and is the single most effective thing you can do. Put something soft behind the microphone and something soft behind you. Record a clean thirty seconds of silence at the start of every session, so that if you do need a noise profile later, you have a good one.
None of this is what anyone wants to read when they already have a problematic file. But the file you make tomorrow will be better for it, and the tools work far better on recordings that needed them less.
Keep reading