Hiss, hum, fan, air conditioning, the sound of the room. We measure how loud yours actually is before touching it — and afterwards we tell you, in decibels, how much of it went.
Drag & drop audio or video, or click to browse
MP3, WAV, FLAC, M4A, OGG, MP4, MOV, MKV, WEBM and anything else ffmpeg reads — up to 500 MB, 2 hours
Upload a recording and get it back with the constant background noise taken out — hiss from a cheap microphone, hum from the mains, the flat roar of a fan or an air conditioner, the sound of a room. The work is done with ffmpeg on our server rather than in your browser, so the file can be as large as your connection can carry.
A spectral denoiser needs to be told one thing above all others: how loud the noise is. Get that wrong and it either leaves the hiss alone or chews through the signal, and no amount of turning the strength up fixes either. Every online denoiser we can find ships a fixed value for it. We measure it from your file — the tenth percentile of fifty-millisecond window energies, which is the level the quiet parts actually sit at — and set the filter from that. Benchmarked here against a known-clean reference, with white, pink and brown noise at three levels each:
| Case | With a fixed setting | Measured from the file |
|---|---|---|
| Loud hiss (5 dB SNR) | 8.57 dB | 17.59 dB |
| Across all nine cases | — | +2.09 dB better on average, never worse |
| Against a sweep of every possible setting | — | within 1.1 dB of the best one |
The loud case is the one that matters, because it is the recording people actually bring to a noise reducer. A file whose hiss is already faint barely needs the tool; a file drowning in it is where a hardcoded setting gives up, and that is exactly where measuring is worth nine decibels.
With the noise floor set correctly, we swept the reduction amount from 6 to 97 and measured the result each time: 6 scored 16.43, 10 scored 19.26, 14 scored 21.12, 20 scored 22.31, and 28 scored 22.61. From 28 to 97 the measurement does not move — 22.62, 22.61, 22.61. A tool that offers you a slider from 0 to 100 is offering seventy-two positions that change nothing, and inviting you to believe the file got cleaner each time you dragged it. This page offers gentle, normal and strong, which are 10, 20 and 28, and that is the whole useful range.
The second engine is a recurrent neural network trained on speech (rnnoise), and it is not simply better. It leaves the silence between words dramatically cleaner — on the same file, a floor of −62 dBFS against the spectral filter's −51 — and on badly noisy speech it is far ahead, because it recognises voice rather than subtracting an average. But on a recording that is already reasonably clean it scored 14.90 against the spectral filter's 31.59, because at that point it is reconstructing speech rather than removing noise, and the reconstruction is worse than the original. So it is offered for the case it wins and it is not the default. It is also the wrong tool for music, where everything that is not a voice is the point.
Noise lives in the soundtrack, so that is the only stream that goes through the filter. The picture is copied across, and we checked it the way such a claim should be checked: by hashing the video packets in the file that went in and the file that came out. Identical. You lose nothing in the image by cleaning the sound, and you do not wait for a re-encode you did not need.
Ten minutes of audio takes 7.8 seconds with the spectral engine and 14.8 with the neural one — about 77 and 40 times faster than the audio plays. Two more passes read the noise floor before and after, which is what makes the number in your result real rather than promotional; on a long file those are capped at the first three minutes, because the noise floor is a property of the microphone and the room, not of the minute you happen to measure.
Free use is five minutes a day, from files up to 100 MB. A credit pack raises that to 500 MB per file and two hours of audio or video. Because denoising always rewrites the samples, there is no lossless path here the way there is with our trimmer: you get back the format you sent, at a quality where the encoder is not the thing you can hear — 320 kbps MP3, or the original codec where it is lossless. Uploads are deleted as soon as the job finishes and the result six hours later.
The constant kind — anything that is there the whole time at roughly the same level. Microphone hiss, mains hum and its harmonics, fan and air-conditioning roar, computer noise, the general tone of a room. That is what a spectral denoiser is: it works out what is always present and subtracts it. It cannot remove a door slamming, a dog barking, a chair scraping or someone talking in the background, because those are events rather than a floor, and nothing that removes them can avoid removing parts of your recording too. If your problem is a single loud bang, a trim is a better tool than a denoiser.
We will not guess, and the page does not: the server measures your file's noise floor before and after, the same way both times, and the result tells you the difference in decibels. On our benchmark files the normal setting dropped the floor by 18 to 20 dB, which is the difference between an audible hiss and a background you stop noticing. But it depends entirely on what you upload, and when a file cannot be measurably improved the result says that instead of dressing up a one-decibel change as a success.
Start with the first one. It suits everything — music, room tone, mixed recordings — and it is the one whose settings we measured your file for. Choose the voice engine when the recording is speech and the noise is bad: it is a speech model, so it clears the gaps between words far better than subtraction can, and on badly noisy speech it was well ahead in our tests. Do not choose it for music, and do not choose it for a recording that is already fairly clean — on a clean file it scored less than half as well as the other engine, because it starts rebuilding speech rather than removing noise.
Yes, and the picture is not touched. Only the soundtrack goes through the filter; the video stream is copied across, which we verified by hashing the video packets of the input and the output and comparing them — identical. You get the same file back with a cleaner soundtrack, in the same container where the format allows it (MP4, MOV, MKV, WEBM) and in MKV where it does not. The audio is re-encoded because it has to be; the picture is not, because it does not.
Some, always — removing noise means changing the signal, and anyone who tells you otherwise is selling something. The failure mode of spectral denoising is a watery, metallic quality on what is left, and it gets worse the harder you push. That is why the strong setting stops at 28 rather than going to 100: past that point our measurements show the result stops improving while the artefacts keep arriving. If the cleaned file sounds strange, try gentle — it removes less and keeps more, and on a lightly noisy recording it is often the better trade.
Audio: MP3, WAV, FLAC, M4A, AAC, OGG, Opus, WMA and anything else ffmpeg reads. Video: MP4, MOV, MKV, WEBM, AVI and the rest. You get back the same format where we can — an MP3 returns an MP3 at 320 kbps, a WAV returns a WAV, a FLAC returns a FLAC — and MP3 for the odd formats nothing else handles well. Denoising rewrites every sample, so unlike our trimmer there is no copy path and the audio is always re-encoded; the settings are chosen so the encoder is not what you hear.
Seconds, for most files. Ten minutes of audio takes about 7.8 seconds with the standard engine and 14.8 with the voice one — roughly 77 and 40 times faster than the recording plays. Two short measuring passes are added on top, which is what lets the result report a real before-and-after rather than a brochure figure. Uploading a large video is usually the slow half, not the cleaning.
Your upload is deleted as soon as the job finishes, and the cleaned file six hours later. That window exists only so the download link still works if you close the tab and come back. Nothing is kept after that, nothing is listened to, and nothing is passed to anyone — which is worth stating plainly for a tool people bring interviews, voice notes and unreleased recordings to.