Removing Background Noise Without Destroying the Voice
Noise reduction is a trade, not a repair. Which noises can be removed cleanly, which cannot, how to hear the artefacts you are introducing, and the order that keeps a voice sounding human.
Updated August 25, 20267 min read
Every noise reduction tool is making the same bargain on your behalf: it removes some of the noise and some of the voice, and you choose the ratio. Understanding that stops you from doing the thing everybody does first, which is to push the slider until the noise is gone and then wonder why the speaker now sounds as though they are transmitting from the bottom of a well.
Steady noise comes out, sudden noise does not
Classic noise reduction works by learning what the noise looks like across the frequency spectrum and subtracting that profile from every moment of the recording. This works beautifully when the noise is constant, because a profile measured during a pause is still accurate three minutes later. It fails on anything intermittent, because there is no stable profile to subtract — and subtracting an average of a sound that comes and goes damages the moments where it was absent. The noise reduction tool has three modes: Noise profile, which learns the noise from a stretch you mark on the waveform; Spectral, which lowers everything under a noise floor you set; and AI Voice, which is trained on speech and treats everything else, music included, as noise — so use Spectral or Noise profile on music.
| Noise | Removable? | Notes |
|---|---|---|
| Air conditioning, fans, computer hum | Yes, very well | Constant and mostly low frequency |
| Mains hum at 50 or 60 Hz | Yes | A narrow notch removes it almost perfectly |
| Traffic rumble, distant motorway | Largely | Steady enough for a good profile |
| Tape hiss, preamp hiss | Yes | Broadband but constant |
| A door slamming, a cough | No | Cut or replace that moment instead |
| Keyboard clicks | Poorly | Too brief and too broadband |
| Another person talking | No | Same frequencies as your speaker |
| Room echo and reverb | No | Not noise — it is your own voice, delayed |
The last two rows are the ones that disappoint people most. Reverb in particular is often described as something to be removed, but it is not an additive layer sitting on top of the voice — it is the voice itself arriving repeatedly from every wall. Nothing separates it out reliably after the fact. The fix is a closer microphone and a softer room, applied next time.
The artefacts you are trading for
Push reduction too far and the recording acquires a characteristic set of flaws. They are easy to hear once you know what to listen for, and nearly invisible until then — which is why so much amateur audio arrives over-processed.
- The underwater voice. The high frequencies that carry consonants have been removed along with the hiss, so *s* and *t* sounds go soft and the whole thing feels muffled and distant.
- Musical noise. A faint burbling or ringing in the gaps between words, like tiny bells. This is the classic signature of aggressive spectral subtraction and there is no mistaking it once heard.
- Breathing that cuts in and out. A noise gate opening and closing on quiet passages, so the room tone appears and disappears abruptly around each phrase.
- Pumping. The background level audibly rising and falling in time with the speech, which draws attention to the processing rather than hiding it.
- Truncated word endings. Trailing consonants clipped off because they fell below the threshold, making speech sound clipped and slightly rude.
The practical guidance is to remove less than you want to. A little residual noise is unremarkable to a listener — every professional recording has a noise floor — whereas processing artefacts are immediately noticeable because nothing in nature sounds like them. Aim to make the noise unobtrusive rather than absent.
Order of operations
Cleaning up a voice recording is a short sequence, and doing it out of order either wastes effort or bakes in problems. Each step here makes the next one easier.
Trim first
Cut the recording down to what you are actually keeping before processing anything. There is no point denoising four minutes of setup chatter, and a shorter file makes every subsequent step faster.
High-pass to remove rumble
In the equaliser, roll off everything below about 80 Hz for a typical voice. Nothing useful in speech lives down there, but traffic, air handling, desk bumps and microphone handling all do. This one move often removes half the perceived noise before any denoising happens.
Denoise, gently
Apply noise reduction at a moderate setting and listen to the gaps. If you hear burbling, reduce it. Two light passes cause fewer artefacts than one heavy pass, so it is worth repeating rather than escalating.
Deal with individual noises by hand
A single cough or slammed door is an editing problem, not a processing one. Cut it, or replace that fraction of a second with room tone copied from a pause, which is inaudible where a silent gap would be conspicuous.
Level it last
Run the normaliser at the end, once the noise floor is where you want it. Normalising first raises the noise along with the voice and gives the denoiser a harder job.
Different jobs, different targets
| Recording | How hard to process | Why |
|---|---|---|
| Podcast or interview | Moderate | Listeners tolerate some room tone; artefacts over an hour are exhausting |
| Voiceover for video | Moderate to firm | Music and effects will mask residual noise anyway |
| Audio for transcription | Light | Recognisers cope with noise better than with artefacts |
| Music recording | Very light | Denoising takes air and reverb tails with it |
| Evidence or archive | None | Process a copy; keep the original untouched |
| Phone call recording | Firm | Already heavily compressed — little left to protect |
The transcription row surprises people. It is tempting to scrub a recording hard before running speech to text, but recognition models were trained on ordinary imperfect audio and handle background noise well, while the artefacts of heavy denoising are unlike anything in their training data. A light touch produces a better transcript than a thorough one.
Reduce background noise on your device — the recording is never uploaded.
Clean up a recordingPreventing it at the source
Everything above is repair work, and repair work has a ceiling. Four changes at record time each do more than any processing chain, and none of them requires buying anything.
- Get closer to the microphone. Halving the distance roughly quadruples the voice level relative to the room. A hand's width away is a different recording from an arm's length away.
- Turn things off. Air conditioning, fans, fridges, notifications. Thirty seconds of walking round the room removes noise that no software fully removes afterwards.
- Soften the room. Curtains, a rug, a bookshelf, a wardrobe of clothes. Bare walls and a hard desk are what create the reverb you cannot fix later.
- Record thirty seconds of silence. Deliberately capture the room with nobody speaking, at the start of every session. It gives noise reduction a clean profile and gives you material to patch over edits.
A word on where this processing happens. Audio recordings are frequently confidential — interviews, consultations, internal meetings, unreleased music — and the usual online denoiser uploads the file to a server to process it. These tools run on your device, not our servers, so there is no upload wait on a long file and no copy is left on a server after you close the page. The free plan takes one file up to 50 MB and 15 minutes; Pro removes those caps and takes up to 25 files at once.
Frequently asked questions
How do I remove background noise from a voice recording?
Trim to what you are keeping, roll off everything below about 80 Hz to remove rumble, then apply a moderate amount of noise reduction and check the pauses for burbling artefacts. Level the result last. Two light passes cause fewer artefacts than one aggressive pass.
Can I remove echo from a recording?
Not reliably. Echo and reverb are the same voice arriving repeatedly after bouncing off surfaces, not a separate layer that can be subtracted, so tools that claim to remove it are guessing. The real fix is recording closer to the microphone in a room with soft furnishings.
Why does my audio sound muffled after noise reduction?
Because the reduction was too strong and took the high frequencies with it. Hiss and consonants occupy overlapping parts of the spectrum, so removing all the hiss also removes the s and t sounds that make speech intelligible. Lower the setting until a little noise remains.
Should I denoise audio before transcribing it?
Only lightly. Speech recognition handles ordinary background noise well, but processing artefacts are unlike anything in its training data and hurt accuracy more than the noise did. Remove an obvious constant hum, then stop.