How to remove background noise from a recording
Fan hum, room tone, traffic and keyboard clatter each need a different fix. What neural noise suppression does well, what it cannot fix, and how to record so you need less of it.
Updated 2026-09-19
Noise removal has improved enormously in the last few years. Tools built on small neural networks can strip a constant hum or air-conditioner rumble from speech almost perfectly, in real time, in a browser tab. But no algorithm can put back what the noise destroyed, and the wrong settings turn a voice into an underwater robot. Knowing the categories of noise tells you what to expect.
Four kinds of noise, four outcomes
Steady broadband noise — fans, air conditioning, tape hiss, electrical hum. This is the easy case. Neural suppressors such as RNNoise learn what speech looks like and treat everything else as noise; a steady bed of hiss disappears with almost no damage to the voice.
Intermittent noise — keyboard clicks, doors, a dog barking. Harder. Modern suppressors catch most short transients between words but struggle when the click overlaps a consonant. Expect a big improvement, not perfection.
Other voices and music. Suppressors tuned for speech will keep other speech; they cannot tell your voice from a colleague's. Music is partly removed, partly smeared. If this is your problem you need source separation, which is a different tool.
Clipping and distortion. Not noise at all — the waveform was cut off during recording. No suppressor fixes this; de-clip tools can reconstruct mild cases.
How neural suppression works, briefly
RNNoise, released by the Xiph.Org Foundation, splits audio into short frames, estimates for each frequency band how much of the energy is speech, and scales the noise bands down. The network is small enough to run faster than real time on a phone, which is why it can run locally in a browser without sending your recording anywhere.
Strength is a mix, not a switch
Full-strength suppression on a mildly noisy recording sounds processed: sibilants soften and room ambience vanishes entirely, which listeners perceive as unnatural. Blend the cleaned signal with the original — 60 to 80 percent cleaned is often more pleasant than 100. Listen on headphones and stop when the noise stops bothering you, not when the meter reads zero.
Record so you need less of it
Get the microphone closer: halving the distance gives you roughly 6 dB more voice for the same noise. Turn off fans for the take. Record a few seconds of silence at the start — some tools use it as a noise profile, and it tells you how bad the floor is before you commit an hour.
Step by step
- 1
Identify the noise
Listen to a gap between words. Steady hiss or hum: proceed. Clicks and voices: lower expectations. Distortion: use a de-clipper instead.
- 2
Denoise a short section first
Process 20–30 seconds, compare with the original at matched volume, and pick a strength.
- 3
Process the whole file
Apply the chosen strength once. Avoid running suppression twice; artefacts compound.
- 4
Level and export
Cleaned speech usually sounds quieter. Normalise loudness afterwards and export at 128–192 kbps for speech.
Questions
- Will noise removal make my voice sound robotic?
- At full strength on a very noisy file, possibly. Reduce the strength or mix in some of the original signal.
- Can I remove music from behind a voice?
- Speech-focused suppressors reduce it but leave artefacts. Use a dedicated vocal/instrument separator for that job.
- Is it safe to denoise a confidential recording online?
- Only with a tool that processes in your browser and never uploads. Check for an explicit statement; if the site shows an upload progress bar, the file is leaving your device.