Remove background noise from a recording
A recording from the street, from a café or from a lecture hall carries the noise of the place into every word. This tool takes that noise down with a high-pass filter, a noise gate that opens on the voice and closes over the pauses, and a compressor — in your browser, on your own device. The file is never uploaded.
The recording
Drop the recording here
or , or
No recording yet.
The strength sets how close the gate comes to the measured noise floor. At 0 % the recording is left exactly as it is.
Choose a recording to start.
The recording is decoded by the Web Audio API of your browser and cleaned by JavaScript on your own device. Nothing is uploaded, the page loads no third-party code, and there is no form and no account.
What changed
- Noise floor before
- —
- Noise floor after
- —
- Difference
- —
- Length
- —
- Sample rate
- —
- Channels
- —
The numbers appear as soon as a recording is loaded. They are measured on your file, not taken from a table.
What this tool is — and what it is not. It is a noise gate and filters: a high-pass filter, a noise gate and a compressor, written by hand in JavaScript. It is not a neural network and it does not restore a voice the way a studio or a dedicated restoration service can. It lowers the measured noise floor of your recording, and that is exactly what the panel on the left shows.
How to clean a recording before you transcribe it
-
Open the recording
Drop the file on the box, choose it with the button, or try the sample of the site. The browser decodes it on your own device.
-
Set the strength
Start at 60 %. The gate follows the measured noise floor of your file, so a recording with a loud hum needs more than a quiet room.
-
Listen before and after
The two buttons play the same recording, cleaned and untouched, so you hear the difference instead of believing a number.
-
Download the WAV
The cleaned recording is written as a 16-bit WAV file on your own disk. Nothing was uploaded to get there.
-
Send it to the transcription
Need the text and not only the sound? Upload the cleaned file to Transcrib Si and get the transcript with timecodes and speaker labels.
How the noise floor is measured
The panel shows real measurements of your file. Here is what every number means.
| The number | What it is |
|---|---|
| Noise floor | The recording is cut into windows of 20 ms, every window gets its own level, and the quietest tenth of those levels is called the floor. It is written in dBFS, the scale where 0 dB is the full scale of the recording. |
| Difference | How far the floor came down after the cleaning. A negative number means the floor is lower than it was — that is the result of the cleaning, measured on your file. |
| Length, sample rate, channels | What the browser read from the file. The cleaning does not change any of the three, so the cleaned recording is the same recording. |
| The strength | It sets the corner of the high-pass filter, how close the gate comes to the measured floor, how hard the gate works below it, and how much the compressor turns the loud peaks down. At 0 % not a single sample is touched. |
The measurement is a property of your recording and of the strength you picked. It is not a score, and it does not promise the same number for another file.
What the three stages do
- The high-pass filter. Everything below 70–100 Hz — the rumble of a bus, of a fridge, of a table that is touched — is turned down. The voice band above it is left alone.
- The noise gate. The tool measures the level of every tenth of a second and compares it with the measured floor. Where the voice is, the gate opens at once; over the pauses it closes, and the hum between the words and the traffic that does not stop go down with it. The gate works on both channels at once, so the stereo image does not swim.
- The guard of the gate. If the recording has no quiet parts at all — music, a held note, a file of pure noise — there is nothing the gate could keep and something it could throw away, so the gate stays open and leaves the recording alone. It never closes on the sound itself.
- The compressor. The loud peaks are turned down towards the rest, so the recording is easier to hear and easier to transcribe. The level of the recording is never raised, so the tool cannot add noise of its own.
Where this tool stops. A gate cannot bring back a voice that was recorded badly, and it cannot separate two people who speak at the same time. If the recording is damaged in that way, the honest answer is a better microphone, a better place, or a service that restores the sound — not a browser filter.
Questions this page answers
Does this tool really run in my browser?
Yes. The recording is decoded by the Web Audio API of your browser, and the cleaning is done by JavaScript that runs on your own device. Nothing is uploaded, the page loads no third-party code, there is no form and no account, and nothing is written to a log. The server only sends the page itself, the same way it does for every visitor.
What can the tool actually do?
Three things, and it is honest about them: a high-pass filter that takes the rumble under 70–100 Hz down, a noise gate that opens on the voice and closes over the pauses, taking down what stays below the measured noise floor, and a compressor that turns the loud peaks down. It is not a neural network and it does not restore a damaged voice — it lowers the noise floor of your recording, and the panel shows the measured result. If the file has no quiet parts at all, the gate stays open and leaves it alone.
Is my recording safe?
The file stays on your device. It is read with the File API, decoded in the memory of the browser and cleaned by JavaScript in the same tab. There is no upload, no account and no third-party code on the page. The only thing that leaves your device is the cleaned file, and only when you press the download button.
How do I save the cleaned recording?
Press “Download WAV”. The tool writes a 16-bit PCM WAV file with the same length, the same sample rate and the same number of channels as the recording you loaded, and the browser saves it to your downloads as a .wav file.
Related pages
Need the text and not only the sound? Send the cleaned file to the transcription service.
What Transcrib Si is and what it costs.
The other free browser tool of the site: a SubRip file becomes a WebVTT file.
The same, the other way round.
How long a text takes to say out loud, at three measured paces.
What happens to a file that you send to the service.
Mark Burmistrov, Founder.
The tool is written by hand: a high-pass filter, a noise gate and a compressor, with the measurement of the noise floor taken from your own file and shown to you. The author of the page and the publisher of the site are the same person. Questions and corrections: support@transcrib.si.