Removing background noise from audio: what works, and what cannot be saved

The file is sitting there. The take is good, the read is right, and underneath it there is a hiss that will not go away, or a door closing exactly under the best line. The instinct is to open a clean-up tool and push the slider until the noise stops. That is usually the precise moment a recording starts getting worse instead of better.

You think you are looking for a treatment. What you need first is a diagnosis.

Which noise are you actually dealing with?

Four families cover almost everything that gets filed under background noise. They do not share a cause, they do not share an answer, and this is the step nearly everyone skips.

• Constant hiss. Ventilation, air conditioning, a computer fan, a preamp pushed too hard. It runs from the first second to the last, stable, spread across a wide band of frequencies, part of which sits directly on top of the voice.
• One-off events. A door, a notification, a passing vehicle, a keyboard, a dog. Each has a beginning and an end, and spoils only a few moments of the file.
• Room reverberation. Nothing was added to the voice here: the voice itself comes back to the microphone, reflected by bare walls, with a tiny delay. You hear it as a hollow, distant sound, the bathroom effect.
• Electrical hum. A low, very regular drone, usually traced back to an earth connection, a cable, a power supply, or a light fitting sitting close to the microphone.

To settle which one you have, put headphones on and listen to the stretches where nobody speaks, between two lines and before the first. That is where noise shows itself undressed. A hiss identical from the first silence to the last, a drone that pulses, an isolated crack: a few attentive listens teach you more than an afternoon of settings.

(Clipping baked in at the recording stage, a level set badly, plosives that were never filtered: those are faults of a different nature, they do not answer to noise treatment, and diagnosing them is a subject of its own.)

What actually works, family by family

There is no single way to remove background noise from audio, because there is no single background noise. Most of the advice available starts at the slider and stops there. What follows is method, so the reasoning holds whatever you happen to have open.

Constant hiss: start from a fingerprint

Serious noise reduction begins with a sample. You isolate a passage of real silence, with no breath and no page turn in it, and you let the treatment learn what it must fight. That fingerprint has to come from the same session, the same microphone, the same position and the same gain as the voice you want to clean. A fingerprint borrowed from elsewhere describes a noise that is not yours, and the treatment goes hunting for things that were never in the file.

After that, one rule holds: small increments. Two light passes almost always beat one aggressive pass. Come down gradually, and listen to the voice rather than to the silence.

One-off events: edit rather than process

Running a global reduction to erase a door slam punishes the entire file for a one-second offense. Isolate the region instead. If the noise falls between two words, cut it and drop in a piece of silence lifted from the same session, so the floor stays continuous to the ear. If it falls underneath a word, you arbitrate: record the line again, or accept it. There is no honest third option.

Hum: look for the cause before reaching for a filter

Electrical hum repeats itself, which is what makes it testable. Change the cable, move the power supply away, unplug the laptop charger, take the microphone away from a lamp or a screen, then record a few seconds again. If the drone goes, you have solved it at source, and no treatment does that as well. If it survives, it is the only noise on this list that comes out surgically: it occupies a narrow frequency and its multiples, and you can work on it without harming the rest.

The frequencies that carry no voice at all

Below the fundamental of a speaking voice there is nothing useful to keep: traffic rumble, footsteps on the floor above, vibration traveling up the microphone stand, moving air. Clearing that region lightens a recording immediately, and it does so without touching intelligibility.

No universal value works here. A deep male voice and a bright female voice do not sit on the same floor. The method: raise the cut-off point until you hear the voice thin out, then come back down until the body returns. Your ear settles this, and no figure read somewhere else can settle it for you.

The point nobody makes: noise reduction is subtractive

Every denoising treatment works by taking energy out of the signal. Noise and voice share frequencies. Take one away and you always take a little of the other with it. Below a certain threshold that loss is imperceptible. Above it, you hear it, and it does not repair.

The objection is fair: the tools have improved a great deal, and some do remarkable work on material that would once have been written off. True. It does not change the arithmetic underneath.

Here is what tells you that you have gone too far.

• A metallic shimmer around the words, usually described as an underwater voice, or a slightly robotic one.
• Consonants eaten away: the s, the f and the t lose their edge, and the diction goes soft.
• Breaths that vanish abruptly and leave flat, unnaturally dead gaps.
• A voice that seems to have stepped backwards, as if the performer had drifted away from the microphone mid-take.

The most reliable test, when the tool allows it, is to listen only to what has been removed. If you recognize fragments of voice in that residue, you are past the balance point. Beyond it you are no longer removing noise, you are removing the voice.

One arbitration gets forgotten almost every time: treat for the destination. A light residual hiss underneath a music bed is inaudible on air. A voice damaged by over-processing stays damaged in the final mix.

What can never be recovered

This needs saying plainly, because it is rarely written down.

A very reverberant room is lost. The reflections are the voice itself, mixed with a delayed copy of itself, and no treatment separates the two cleanly. Attenuation is the ceiling.

Clipping baked in at the recording stage is lost too, for a simpler reason: the information was never captured, so there is nothing to bring back.

A microphone placed too far from the performer froze the ratio between voice and room at the moment of the take. Reducing the room then means reducing the voice with it.

And a voice buried under a source as loud as itself, an extractor fan, a crowd, a machine, offers neither a usable silence for the fingerprint nor a wide enough gap for separation.

In these four cases, no tool and no amount of time change the outcome. The take already decided.

The real question arrives before the record button

Rescue work has genuine value. It saves the one-off interview, the testimony nobody will give twice, the capture that exists in a single copy. What it never does is make a take better than it would have been, clean, from the start. Almost everything that reduces background noise in a recording happens before the record button.

When the content carries a brand name and goes out into distribution, the arbitration moves upstream: a treated room, a microphone distance that is held, a level that keeps headroom, and that reference silence recorded before the first line. For 15 years, that is what our studio in Aix-en-Provence has put in place before the performer opens their mouth.

If the file you are trying to save has to carry your brand, let’s talk about where it will be heard before we talk about cleaning it. Describe your project on our voice over agency page, or send us the details through the quote form.