How to Remove Background Noise From Audio (And What It Cannot Fix)

Most advice about removing background noise amounts to move a slider until the noise goes away. Follow that and you will end up with a recording where the noise has gone and the voice sounds like it is being transmitted from underwater. Noise reduction is a trade, not a repair: everything it removes comes out of the same signal your voice is in, and the skill is in knowing how much to give up.

This guide covers how it actually works, how far to push it, what order to process in, and the four common problems it cannot fix no matter how high you set it.

Not all noise is the same

The single most useful distinction is between noise that is constant and noise that happens once.

• Steady noise stays roughly the same the whole way through: air conditioning, a computer fan, mains hum, tape hiss, distant traffic, the hum of a fridge. This is what noise reduction removes well.
• One-off sounds happen and then stop: a door slamming, a cough, a chair creaking, a notification chime, a car horn. Noise reduction does essentially nothing to these.

How the reduction control behaves

The noise remover on this site has one control, from 0 to 100 per cent, with 50 as the default. It is important to understand what that number means: it is how much of the detected noise is subtracted, not a quality setting. Higher is not better. At 100 per cent the process removes everything it believes is noise, and a large part of what it believes is noise is the quieter detail of your voice.

How to do it, step by step

1. Upload your recording. MP3, WAV, OGG, FLAC, AAC and M4A all work, up to 500 MB.
2. Set the reduction to around 40 to 50 per cent to begin with.
3. Export, and listen back on headphones.
4. Listen specifically to the gaps between words, not to the words themselves. Artefacts always appear in the silences first.
5. If noise is still intrusive, raise the amount by ten and try again. Do not jump to 100 to see what happens.

How to tell when you have gone too far

Over-processed audio has a recognisable sound. Listen for:

• A hollow or underwater quality to the voice.
• A metallic edge, sometimes described as robotic or phasey.
• Warbling or fluttering in quiet passages, where the background swims rather than sitting still.
• Consonants losing their edge, so s and t sounds go soft and speech becomes harder to follow.
• Reverb tails that cut off abruptly instead of decaying.

Process in the right order

Order matters more than most people expect, and getting it wrong makes every step harder.

1. Noise reduction first, while the noise floor is still steady and easy to profile.
2. EQ next, to cut any low rumble and boxiness that survived.
3. Compression after that, to even out the loud and quiet passages.
4. Volume last, to bring the finished result up to a usable level.

The critical one is that noise reduction comes before compression. A compressor raises quiet passages, and the quietest passages in a recording are the gaps between words, which is precisely where the noise lives. Compress first and you amplify the hiss, making it both louder and harder to remove cleanly.

What noise reduction cannot fix

Echo and room reverb

Echo is your own voice arriving a fraction of a second late after bouncing off a wall. It is not background noise; it is a copy of the signal. There is no constant profile to subtract, so noise reduction has nothing to work with. This has to be prevented when recording: get closer to the microphone, add soft furnishings, or record in a smaller and more cluttered room. Bare rooms with hard floors are the worst case.

Music behind speech

Music changes constantly, so it never settles into a steady profile. Noise reduction will not separate a voice from a backing track. Taking a vocal out of a song is a different technique entirely, and it works on the stereo image rather than the noise floor.

Clipping and distortion

If the recording was too loud going in, the waveform has flat tops where the peaks should be, and that information is gone. Noise reduction cannot rebuild it, and neither can anything else. Lower your input level and record again.

Wind on a microphone

Wind noise is broadband, erratic, and loud, often overwhelming the voice completely. Mild wind rumble can be reduced by cutting the lowest frequencies, but heavy wind buffeting is not recoverable. A foam windscreen costs very little and solves it at the source.

Using EQ alongside it

A lot of what people call noise is not a noise floor at all but low-frequency rumble, and EQ removes it more cleanly than noise reduction does. Cutting the 32 Hz and 64 Hz bands hard takes out desk vibration, footsteps, traffic rumble and handling noise, and on speech it costs you nothing, because there is no useful voice content down there.

While you are in there, a cut of three to six decibels around 125 to 250 Hz clears the boxy, muddy quality that small rooms add. A small cut in the right place is almost always more effective than a large boost somewhere else, because boosting raises the noise sitting in that band along with the voice.

The fix that beats all of this

Every minute spent on the recording side saves ten on the processing side, and no amount of processing gets you back to what a clean take would have sounded like. If you can control the recording at all:

• Get the microphone closer. Halving the distance is the single biggest improvement available to you, because it raises your voice relative to everything else in the room.
• Turn off the obvious sources. Fans, air conditioning, dehumidifiers, and anything with a compressor in it.
• Record somewhere soft. A room with curtains, carpet, a sofa and full bookshelves behaves far better than a hard, empty one.
• Record a few seconds of silence at the start. It gives any noise reduction tool a clean profile of the room.
• Watch your levels. Aim for peaks well below the top of the meter, since a quiet clean take is fixable and a clipped one is not.

The short version

• Noise reduction removes steady sound, not one-off sounds.
• Start at 40 to 50 per cent, and judge the result by the gaps between words.
• Hollow, watery or metallic means you have gone too far. Back off.
• Process in the order noise, EQ, compression, volume.
• Echo, background music, clipping and heavy wind are not fixable afterwards.
• A closer microphone beats every tool on this list.