How to Remove Background Noise from Audio (Hiss, Hum, Room and Wind)

To remove background noise, first identify what kind it is, then use the matching tool: a high-pass or notch filter for rumble and hum, spectral noise reduction for steady hiss, and a neural denoiser for changing noise like traffic, keyboards or voices in the next room. Apply the least reduction that makes the noise stop distracting; pushing any denoiser to silence makes the wanted sound worse. And if you can still re-record, fixing the noise at the source beats every plug-in.

For a quick clean-up, SoundGen's free noise reducer runs in your browser.

First: fix it before you record

Noise reduction always trades some of the wanted sound for less noise. The cheapest decibels are the ones you never recorded:

• Get closer to the source. Halving the distance to the mic raises the voice or instrument by about 6 dB relative to the room. It's the biggest single improvement there is.
• Turn things off. Fridges, fans, air conditioning and computer fans are the usual culprits.
• Treat the space, even cheaply. A closet full of clothes, a duvet behind the mic or a sofa-filled room is less reflective than a bare kitchen.
• Set levels properly. Record loud enough that the voice sits well above the noise floor, but with headroom so peaks don't clip.
• Use a windscreen outdoors and a pop filter indoors.
• Hum? Check for ground loops: plug everything into the same power strip, use balanced cables, keep audio cables away from power supplies.
• Record 10 seconds of silence — the room with nobody speaking. Many noise reducers learn from a "noise print", and this gives them a clean one.

Know your noise

Noise	What it sounds like	Character	Best first tool
Hiss	Steady "shhh"	Broadband, constant	Spectral noise reduction
Hum / buzz	Low drone at 50 or 60 Hz plus harmonics	Tonal, constant	Notch filters, hum remover
Rumble	Low thump, traffic, handling noise	Low-frequency	High-pass filter (around 80 Hz for voice)
Room tone / HVAC	Air, fan, fridge	Mostly constant, some tone	Spectral or neural
Wind	Bursts of low-frequency blasting	Changing, low-heavy	High-pass, then neural
Intermittent noise	Keyboards, dogs, traffic, chatter	Changing	Neural denoiser
Reverb / echo	Hollow, distant sound	Part of the signal	Not really "noise" — needs a de-reverb tool

Hum is 50 Hz in most of Europe, Africa, Asia and Australia, and 60 Hz in North America and parts of South America and Asia. Its harmonics (100/150 Hz or 120/180 Hz and up) are why a single notch rarely kills it.

The three families of noise reduction

1. Gates and expanders

A gate turns the signal down whenever it falls below a threshold — between words, between phrases. It's simple and transparent when the noise is quiet, but it does nothing while the person is talking: the noise is still under every word, so a heavy gate makes the noise pump in and out with the speech. Use a gate or expander gently, with a few dB of reduction rather than full silence.

2. Spectral noise reduction (spectral subtraction / spectral gating)

These tools split the audio into many narrow frequency bands, estimate how loud the noise is in each band (from a noise print, or by tracking the quietest moments), and turn each band down when it's close to the noise level. Because it works band by band, it can remove hiss during speech, not just between words.

It works best on steady noise. Its signature failure is "musical noise" — random, watery, bell-like chirps where the noise estimate is wrong — plus a dull, lispy voice when pushed too far. Fixes: less reduction, more smoothing, a better noise print.

3. Neural denoisers

Neural models are trained on huge amounts of clean speech mixed with real-world noise, and learn to output just the speech. DeepFilterNet is a well-known open-source example. Because they've learned what a voice is rather than what the noise is, they cope with changing noise — traffic, keyboards, a TV in another room — that defeats spectral tools.

Their limits: most are trained on speech, so on music they can dull or remove instruments they mistake for noise; they can make breaths and quiet consonants sound clipped; and pushed to the maximum, voices can take on a slightly synthetic quality. Most let you cap the reduction — leaving a little room tone usually sounds more natural than total silence.

A sensible clean-up order

1. High-pass filter to remove rumble below the voice or instrument.
2. Notch out hum if there is any.
3. Denoise: spectral for steady hiss, neural for everything else. Start with modest reduction (around 10–12 dB) and raise it only if needed.
4. Gentle gate or expander if the gaps between phrases still bother you.
5. Then compress and EQ. Compressing first brings the noise up with the quiet parts.

Always A/B against the original at the same loudness. It's easy to fall in love with silence and miss that the voice now sounds underwater.

Doing it in SoundGen

• One file, quickly: the free noise reducer.
• In the studio: right-click an audio clip and choose AI Denoise…. It runs a neural speech-enhancement model based on DeepFilterNet3 entirely in your browser — the audio isn't uploaded, and it costs no credits. Set Max noise reduction (3–60 dB, or unlimited) and a wet/dry Mix, and undo restores the original if you don't like it. It's built for vocals, dialogue and podcasts; on full music mixes use a low reduction or a partial mix.
• Live, adjustable: put the Noise Removal effect (a spectral gate) on a track for steady hiss you want to tweak during the mix.

FAQ

What's the best way to remove background noise from a voice recording? A neural denoiser, after a high-pass filter. It handles both steady and changing noise. For steady hiss only, spectral noise reduction works well too.

Can I remove noise from music, not speech? Carefully. Speech-trained neural models can eat instruments. Use spectral noise reduction with a noise print, at low settings, or separate the stems first and denoise only the part that needs it.

How do I remove hum? Notch filters at 50 or 60 Hz and their harmonics, or a dedicated hum remover. Then fix the cause — usually a ground loop.

Why does my audio sound watery or robotic after noise reduction? Too much reduction. Back it off, give the tool a better noise print, or let a little noise through with a wet/dry mix.

Can noise reduction remove echo? Not really. Echo and reverb are reflections of the wanted sound, not separate noise. Dedicated de-reverb tools exist, but recording in a less reflective space is far more effective.