Account

What new tool do you want?

Video & Audio

Audio Noise Remover (Reduce Background Noise)

Remove hiss, hum and background noise from voice recordings, podcasts and videos' audio — adjustable strength, instant preview and download — processed privately in your browser.

How to use

  1. 1Drop an audio file (or a video to clean its sound).
  2. 2Adjust the noise reduction strength and preview.
  3. 3Download the cleaned audio.

Frequently asked questions

What kind of noise does it remove?

Steady background noise — hiss, hum, fans, air-conditioning and road rumble. Sudden sounds like dogs barking aren't removed.

Why does my voice sound watery?

The strength is too high — lower it until the voice sounds natural.

Is my recording uploaded?

No — it's processed entirely in your browser.

Why the tool reports an exact dB number, not just 'noise removed'

Before and after processing, the tool measures a 'noise floor' estimate: it splits the audio into 50-millisecond windows, computes the loudness (RMS) of each one, sorts them, and averages just the quietest 10% — the windows most likely to be gaps between speech or notes, where only background noise is present. Comparing that measurement before and after denoising gives a genuine, specific number (e.g. 'reduced by about 14 dB') rather than a vague claim, which is a real measurement of how much quieter the background got, not a cosmetic statistic.

Why a recording with no pauses denoises worse than one with a few seconds of silence

The denoising algorithm needs to learn what the background noise actually sounds like before it can subtract it from the signal, and the clearest source of that information is a stretch of audio containing noise alone — a brief pause between sentences, a quiet moment before someone starts talking. A continuous recording with zero quiet moments gives the algorithm nothing to learn the noise profile from, which is exactly why the UI's own guidance flags needing a second or two of pauses: it's not a stylistic suggestion, it's a structural requirement of how noise-profile-based denoising works.

Why turning the strength up too far makes voices sound 'watery'

Noise reduction works by identifying and suppressing frequency content that matches the learned noise profile — but a human voice and steady background noise (hiss, fan hum) inevitably overlap somewhat in frequency, so aggressively suppressing the noise band also shaves off some of the voice's own frequency content, particularly its subtler harmonics. At moderate strength (the default sits at 65%) this tradeoff is barely audible, but pushed toward 100% the algorithm starts removing more of the voice's natural texture along with the noise, producing the hollow, underwater-sounding artifact the FAQ calls 'watery' — which is why the fix is always to dial the strength back, not to apply it twice.

The decibel note and the noise that gets removed are measured on different rulers

Files are read at 44,100 Hz. The reduction itself listens in slices of about 46 milliseconds, stepping forward about 12 milliseconds, and learns the hiss from the quietest 15% of those slices — always at least five of them, even on a very short clip. The line that says the background was reduced by about so many dB uses a different ruler: 50-millisecond windows, only the quietest 10%, turned into decibels, then the before-and-after difference. That number is rounded to a whole decibel, and it is never shown as negative, so a result that got slightly noisier still reads as about 0 dB. The slider opens at 65% and runs from 10% to 100%. Left and right are cleaned separately.

At 65%, the cut goes past the measured noise, but a trace of every frequency stays

How far past the measured noise the cut reaches is 1 plus 2.5 times the strength. At the default 65% that is about 2.6, so the tool subtracts more than the noise it estimated. The quietest any frequency is allowed to become is 1 minus 0.92 times the strength, about 40% of its original level at 65%. At 100% that floor is 8% and the cut is 3.5 times the estimate. At the bottom of the slider, 10%, the cut is only 1.25 times and about 91% of each frequency is kept. Even at full strength a little of every frequency remains, which is why a voice turns watery instead of vanishing. Loud moments are turned down at once, and the reduction eases off more slowly when the noise returns.

Related tools

More: Video & Audio