Audio noise remover
Take the hiss, hum, traffic and room echo off a voice recording. A speech model runs on your own machine, on the whole recording, and you hear the before and after at the same moment before you save anything.
Open the recording, run the model, compare and save.
Signed in, typed text and settings save automatically to your private account. Text opened from files stays in this browser’s local storage. Files stay in your browser. Account saving
Drop a recording here, open one, or record
It is read in this tab. Nothing is uploaded, there is no queue, and you get the whole file back rather than a preview.
MP3, WAV, M4A, AAC, FLAC, OGG and Opus, whatever your browser can decode. , or .
Below 100% the cleaned signal is mixed back with the original.
Measured to BS.1770, lifted by at most 9 dB, held under -1.5 dBFS.
Cleaned
Keyboard
| Space | Play or pause |
| B | Switch between the original and the cleaned version |
| ←→ | Skip five seconds |
| Home | Back to the start |
The recording is decoded and processed in this tab. It is never uploaded, and it never enters your account.
How to remove background noise from a recording
Open the recording
Drop an audio file onto the waveform or press Open a file. Anything your browser can decode works, including MP3, WAV, M4A, FLAC and Opus.
Press Remove the noise
The first run downloads the engine, then the model works through the file. Ten minutes of audio is about half a minute of work on a laptop, and longer on a phone.
Compare, then save
Switch between Original and Cleaned to hear the same moment both ways, then save the whole thing as WAV or MP3.
Using the audio noise remover
There are two settings and one button. This is what each control does and when you would reach for it.
Opening a recording
Drop a file onto the waveform, press Open a file, or press Try the example recording to hear what the model does before you commit your own file to it. Decoding is done by your browser, so the formats it takes are the ones your browser already reads.
| What you open | What happens |
|---|---|
| MP3, WAV, M4A, AAC, FLAC, OGG, Opus | Decoded and resampled to 48 kHz, which is the only rate the model runs at |
| A stereo file | Each channel is cleaned on its own, and they share one loudness pass and one limiter, so the mastering never moves the image |
| An MP4 or WebM video | The audio track is decoded and cleaned. You get an audio file back, not a video |
| Longer than 40 minutes | Refused, with the reason. Cut it up in the audio trimmer first |
The waveform, and what the two colours mean
Before the run, the waveform is your recording in grey. After it, the cleaned version is drawn over the top in the accent colour and the original stays behind it. The interesting part is the gaps between phrases: the grey that remains there is the noise floor, and the distance between the two is what the model took out.
Click anywhere on the waveform to move the playhead there. The stage is also the drop target, so you can drag another recording straight onto it when you are done.
Listening to the difference
The Original and Cleaned buttons on the transport switch what you are hearing without losing your place. If it is playing when you press one, the other version starts from the same instant. That is the comparison worth making, and it is the one that tells you whether the strength setting is right.
| Key | What it does |
|---|---|
| Space | Play or pause |
| B | Switch between the original and the cleaned version |
| ← → | Skip five seconds |
| Home | Back to the start |
The keys work when the waveform has focus, so click it or tab to it first. Full screen moves the whole panel into the viewport, which is worth doing on a long recording because the waveform gets the width.
Strength
At 100% you hear the model's output. Below that, the cleaned signal is mixed back with the original, so 70% is seven parts cleaned to three parts what you recorded. The noise comes back in proportion, and so does the room.
Reach for it when the result sounds too processed. A heavily denoised voice can sound like it is in a padded box, and letting a little of the original room back in often sounds more natural than the perfectly clean version. Around 80% is a good place to try first.
Loudness
Removing noise always makes a recording quieter, because noise was part of what you were measuring. The loudness pass measures the result properly and lifts it to a delivery target, rather than normalising to the loudest peak, which on speech is usually a chair creak.
| Setting | Target | When |
|---|---|---|
| Podcast | -19 LUFS | The default. Apple Podcasts and most spoken word platforms |
| Music streaming | -14 LUFS | Spotify and YouTube, and anything going out beside music |
| Broadcast | -23 LUFS | EBU R128 delivery |
| Leave the level alone | none | You are taking this into an editor and will set the level there |
Two limits are deliberate. The lift stops at 9 dB, so a very quiet recording comes back under its target rather than having its remaining noise floor dragged up to meet the voice. The ceiling is held at -1.5 dBFS by a look-ahead limiter rather than by turning the whole file down. The panel reports the loudness it measured, the gain it applied and the peak it finished at, so you can see when the 9 dB cap was the thing that stopped it.
Saving the result
WAV gives you exactly what the model produced, 48 kHz and 16 bit, which is the one to take into an editor. MP3 is encoded here at 192 kbps rather than the usual 128, because a denoised recording is nothing but speech transients and that is precisely what a low bitrate spends its errors on.
Both are the whole recording. There is no preview length, no watermark, no expiry on a download link and no account, because there is no server involved to impose one.
What this tool will not do
It is a speech model, and everything it is bad at follows from that.
- A second voice is not noise. Someone talking in the background is speech, so the model keeps it. This is not a speaker separator.
- Music suffers. A song, or a music bed under a voiceover, comes back thin and strange. Clean the voice before the music goes on, not after.
- Clipping is gone already. If the recording was distorted going in, the information is not in the file and no model invents it back.
- It does not write video. The audio from a video file is cleaned, but putting it back into the video is a job for an editor.
- One file at a time. There is no batch queue. Everything is happening on your own processor, so a batch would just be the same work with less feedback.
- It is not a transcriber. No text comes out. If that is what you want, clean the file here and then open it in audio to text.
What it actually does to a recording
On the example recording on this page, the level in the gaps between phrases falls by 32.5 dB while the speech stays where it was. That is the number the tool exists to produce. The guide beside this page has the rest of them, including a comparison against a noise gate and two other methods on the same recording.
Speed, on an M1 MacBook in Chrome: 65 seconds of audio in 3.35 seconds, which is 19 times real time. A ten minute recording is about half a minute of work. The first run is slower than that because it is also downloading 20.5 MB.
How this compares to the tools that rank for it
Every one of these was run on the same 10.6 second test recording in September 2026. The row that matters is the last one, and floi loses it.
| Tool | Where it runs | What free gets you |
|---|---|---|
| floi | Your browser | The whole file, every time |
| Cleanaudio | Their server | A 30 second preview, then payment |
| Noise Reducer | Their server | A preview, then $0.02 a minute |
| MyEdit | Their server | One download a day, after signing in |
| SimpleClean | Their server | A 30 second preview, then $0.50 up |
| Adobe Podcast | Their server | An account, and the strength controls are paid |
| Where they win: all of them start instantly and none of them makes you download 20.5 MB before the first run. A server with the model already loaded will always beat a cold browser on the first file. | ||
MyEdit's page says its noise remover "runs entirely in your browser", and the same sentence goes on to say "upload your file". It means there is nothing to install. Watching the network while it works shows the audio going to their server, which is worth knowing if you read that line as an answer about privacy.
Frequently asked questions
Is my recording uploaded?
No. The model runs in your browser, so the audio is decoded and processed in the tab and never sent anywhere. The processing engine downloads 20.5 MB the first time you use the tool, after which it is in your browser cache and the tool works with the network switched off. The recording never enters your floi account. Signed-in settings and numeric activity can save privately; Google Analytics receives usage events without audio.
Why do other noise removers only let me hear 30 seconds?
Because their model runs on their server, and server time costs them money. Every one of the tools ranking for this term either caps the free result at a 30 second preview, meters you to one download a day behind a sign-in, or charges by the minute. Here the model runs on your machine, so there is nothing to meter: you get the whole file on the first go.
What kind of noise does it remove?
Steady background that is not speech. Fan and air conditioning hum, mains buzz, computer noise, traffic and street wash, hiss from a cheap microphone or a gained-up preamp, and a fair amount of room reverb. It is a speech model, so it works by deciding what in the recording is a voice and keeping that.
What can it not fix?
Another person talking, because that is speech too and the model keeps it. Music, whether it is the content or a bed under the voice, comes out damaged. Clipping is already lost information and no amount of processing invents it back. And a recording where the noise is louder than the voice comes back clean and thin, because there was very little voice there to keep.
Can I clean the audio in a video file?
Not in one step. Your browser will decode the audio track of an MP4 or a WebM if you drop one in, and you will get cleaned audio back, but this tool writes audio files and cannot put the track back into the video. You would need a video editor for that last step, and the guide next to this page says how.
How long can the recording be?
Forty minutes. The whole file is held in the tab twice over, once as you opened it and once cleaned, and 48 kHz stereo is 384 KB a second, so an hour would be around 2.8 GB before anything is exported. If yours is longer, cut it into pieces in the audio trimmer first.
Does it work on a phone?
It runs, and it is slower. The engine and model are the same 20.5 MB download, which is worth thinking about on mobile data, and the processing is a few times slower than a laptop. For anything longer than a couple of minutes, a computer is the better place to do this.
Which model is this?
DeepFilterNet 3, the published v0.5.6 checkpoint, exported to ONNX and run here with ONNX Runtime Web. It is dual licensed under Apache 2.0 or MIT by Hendrik Schröter and contributors. The signal path around it was written for this tool and checked against the project’s own command line build, which is the measurement on the guide.
Something missing? Request a feature