Video BasicsAug 9, 2026

How to remove background noise from video

Most noise problems are really signal problems. Here is the order to work in, why heavy denoising can sound worse than the noise it removed, and when to leave the floor alone.

8 minute read

Satvik ShrivasFounder, FrameOS

Work out which problem you actually have

Background noise is four different problems wearing the same name, and only one of them is genuinely easy. Steady, unchanging sound — an air conditioner, a fridge compressor, a laptop fan, mains hum — is the tractable case, because it is predictable enough for a processor to model. Intermittent noise like a door, a siren or a chair scrape has to be cut or covered, not filtered, because there is nothing consistent to subtract. Reverb is not noise at all: it is your own voice arriving late off the walls, and no denoiser separates it out cleanly. And a voice recorded from too far away is a signal problem dressed as a noise problem, where the room is not loud so much as the speaker is quiet. Sort your recording into one of those four before you open anything, because they do not share a fix.

Nothing in post is as cheap as microphone distance

In free field conditions, sound pressure drops by about 6 dB for every doubling of distance from the source, which Sweetwater's explainer illustrates with a singer moving from two inches to four and losing 6 dB in the process. Run that backwards and it becomes the most valuable trick in recording: move the microphone from twelve inches to six and your voice comes up by roughly 6 dB, while the fridge across the room stays exactly where it was. You have improved the ratio between the two without processing anything. Real rooms are messier than the theory, since reflections and boundaries flatten the curve, but the direction holds. Close the windows, switch off the fan, mute notifications, record in a smaller room with soft furnishings rather than a big bare one, and get the mic near your mouth. Every one of those is free and permanent.

Record thirty seconds of nothing before you pack up

At the end of a session, keep recording and stop moving for half a minute. That clip of pure background is called room tone, and it is the one asset you cannot manufacture afterwards. Almost every noise reduction tool works by learning what the noise sounds like on its own and then removing that pattern from the rest, so a clean sample of the noise is the difference between a precise repair and a guess. Audacity's documentation is explicit about this: you point it at a section that is only background noise, capture a noise profile, and then apply the reduction. Room tone also patches silences between edits so cuts stop announcing themselves. Same microphone, same gain, same position, same room. Thirty seconds costs you nothing and it is the first thing an editor will ask for.

Hum is the one problem with an exact answer

Electrical hum is not broadband noise. It sits at the frequency of the local mains supply, 60 Hz in the US and 50 Hz across most of Europe, plus a stack of harmonics at multiples of that fundamental. Because the energy is concentrated in a few very narrow bands rather than spread across the spectrum, you can attack it hard without touching much of the voice: a series of narrow notch filters on the fundamental and its harmonics, or a dedicated de-hum module that finds and tracks them for you. This is the one place in this article where aggressive processing is fine. It is also worth fixing the cause, because hum usually comes from a ground loop, a dimmer switch, or a microphone powered off a noisy USB bus, and all three are cheaper to solve than to filter.

Gates and expanders, and why you should leave the floor in

A gate silences everything below a threshold. An expander does the same job proportionally, turning quiet things down rather than off, which is why Sound On Sound describes a gate as an expander with an infinite ratio. For speech an expander is almost always better, because a gate applied to dialogue chatters: the level hovers around the threshold, the gate opens and slams on breaths and word tails, and that draws more attention than the noise did. Hold and hysteresis controls exist to stop exactly that. The more important control is the range or floor, which sets how far the processor drops when it closes. Sound On Sound's advice is to use it to restore a touch of noise rather than muting completely, because the contrast between noisy speech and absolute silence is what the ear notices. Do not aim for silence between words. Aim for the noise not changing.

Spectral repair, and where the tinkly artefacts come from

Spectral noise reduction splits the signal into many frequency bands and pulls down whichever ones are currently sitting at or below the learned noise level. Push it and you get a very specific failure that Audacity's manual names precisely: random bursts of very short tones at random frequencies, described as musical noise, bird song or tinkly bells. It happens when the actual noise floor is higher than the estimate, which in practice means the sensitivity was set too low or the noise profile did not represent the noise across the whole recording. Audacity's own support pages recommend the opposite of what most people do, suggesting less aggressive settings, particularly when reduction is combined with a gate. The frequency smoothing control exists specifically to soften those artefacts. If you cannot hear the noise on headphones at normal listening volume, you are done, whatever the meter says.

AI denoisers are the last resort, not the first move

Modern speech enhancement is genuinely impressive and it is still a filter, not a time machine. iZotope's Dialogue De-noise, for example, runs the signal through 64 psychoacoustically spaced bandpass filters acting as a multiband gate, and its documentation notes that the automatic mode deliberately sets thresholds lower to prevent artefacts, which is an admission that more reduction and more artefacts are the same dial. The whole-file speech models that clean audio in one click go further: they do not subtract noise so much as re-synthesise the voice, which is why heavily processed clips can come back sounding thin, hollow or slightly underwater. Judge them on the awkward moments rather than the clean ones. Listen specifically to laughter, two people talking over each other, the tail of a sentence, and any word said quietly. If those survive, keep it. If they turn glassy, dial it back.

Know when to stop, and when to leave it alone

A steady, low, unchanging noise floor is not a fault. Audiences have listened past hiss for a century. What they do notice is noise that changes: a floor that pumps up and down with the voice, a background that shifts at every cut because each take was processed differently, a room that vanishes into dead silence between sentences. The target is consistency, not purity. Process the whole track with one setting rather than clip by clip, keep the reduction modest, and check the result on the device your audience actually uses, which for short-form means phone speakers with the captions doing half the work. If you are cutting a long recording into a batch of clips, do the audio pass once on the source before clipping, so every clip that comes out of FrameOS or any other tool inherits the same treated track rather than a dozen slightly different ones.

FAQ

How do I remove background noise from a video without it sounding weird?

Use the gentlest tool that works, and stop early. Capture a clean sample of the background, apply modest reduction rather than the maximum, and check the difficult moments — laughter, overlapping speech, quiet word endings. Audacity's manual attributes the tinkly, warbling artefacts people complain about to over-aggressive settings and to a noise profile that does not represent the whole recording.

Can you completely remove background noise from a video?

Rarely, and usually you should not try. Steady noise like hum or fan hiss can be reduced to inaudible. Intermittent sounds and room reverb cannot be filtered out cleanly at all. Total removal also creates dead silence between words, which sounds more artificial than the original noise, so most editors deliberately leave a low, consistent floor in place.

Should I use a noise gate or a noise reduction plugin?

For speech, start with an expander rather than a hard gate, because gates chatter on breaths and word tails. Sound On Sound recommends using the range or floor control to leave a little noise in when the processor closes, so the transition is not jarring. Add spectral noise reduction on top only if a steady floor is still audible, and keep it light.

What is the fastest way to get cleaner audio next time?

Halve the distance between your mouth and the microphone. Sound pressure falls roughly 6 dB per doubling of distance, so moving closer lifts your voice well above a room noise that has not moved at all. Then turn off the obvious offenders, record in a smaller soft room, and capture thirty seconds of room tone before you stop.

Sources

Inverse Square Law · Sweetwater
Production Techniques Using Gates & Expanders · Sound On Sound
Noise Reduction · Audacity Manual
Noise reduction & removal · Audacity Support
Dialogue De-noise · iZotope

Turn one long video into a week of posts

300 credits for 3 days · no card.

Start free