Clean Up Audio

Steady noise that sounds the same all the time: a fan or air conditioner, computer or mains hum, tape or microphone hiss, room tone, distant traffic. The page learns that noise from your noise sample and lowers it wherever the voice does not cover it. Sudden sounds such as a dog barking, keyboard clicks, a door or another person talking change all the time, so they stay; so does echo.

It is classic signal processing, not an AI model. The recording is cut into short overlapping slices of about 40 ms, and each slice into frequency bands. The noise sample gives the average noise level in every band (or, without a sample, the estimate described under "What if no part has background noise alone?"). In each slice, a band that is barely louder than the noise is lowered, by at most the dB you choose, and a band where the voice is clearly louder is left almost as it was. The gain comes from a Wiener filter with a decision-directed estimate (Ephraim and Malah's method), which avoids the warbling "musical noise" of simpler methods. Sensitivity raises the noise estimate, so sounds a little louder than the noise count as noise too. The default of 3 dB leaves an even background when the noise is lowered by up to about 20 dB; beyond that, 6 dB or more stops the leftover noise from chirping, at some cost to the softest parts of the voice.

Everything in the sample counts as noise, so parts of the voice would be removed. The page checks the sample: if its level changes the way speech does, or it is nearly as loud as the loud parts of the recording, it says so and asks for another part. Find the quietest part picks a steady half second for you; where none exists, choose Estimated from the whole recording. After cleaning, listen to Removed noise: it should hold noise only, never words.

Choose Estimated from the whole recording under Noise profile. The page then finds the noise in the recording itself, by minimum statistics (R. Martin, 2001): in every frequency band, the noise is the quietest the recording gets in the 1.8 seconds around each moment, raised by how far such a quietest value lies below the noise's average (measured by the page at your recording's sample rate). A voice comes and goes in every band within that time, even in speech without pauses, while steady noise stays, so the quietest is the noise. It suits steady noise such as a fan, hiss or hum that changes slowly; noise that changes within a second, or a sound held for longer than 1.8 seconds, is less well told apart, and a part of background noise alone, where there is one, gives a more exact profile. The result gives the background's change by the same estimate, and the whole recording's change beside it, so a voice lowered with the noise shows; listen to Removed noise as well.

Loudness in LUFS is measured as ITU-R BS.1770 and EBU R 128 define it: over the whole recording, with silent and very quiet parts left out. −16 LUFS suits podcasts and spoken word, −14 LUFS is the level many music streaming services play at, and −23 LUFS is the European broadcast level. Keep the level changes nothing except peaks above the true-peak ceiling. Raising the level stops in two places: at 24 dB of gain, and where the peaks would need more than 6 dB of limiting, because flattening a voice that far does more harm than the missing loudness. A recording can therefore end below the target, and the result then says which of the two stopped it.

The highest point of the waveform, including peaks between samples, which appear when the file is played or converted to MP3 or AAC. The page measures it with 4× oversampling and keeps it at or under −1 or −2 dBTP. When the loudness target would push a peak higher, a limiter lowers just those peaks, starting 5 ms before them, and the result says by how much.

It removes sound below the frequency you choose with a steep filter (4th order, 24 dB per octave): rumble from handling, traffic or wind, the 50 or 60 Hz tone of mains hum, and any DC offset. 80 Hz suits most voices; a deep voice may sound fuller with 60 Hz. The overtones of hum above that frequency stay unless the noise reduction lowers them.

No. Clipped peaks are flattened in the file itself and cannot be rebuilt; the page counts the clipped places so you know. If you can, record again with a lower input level.

A little. Bands where the voice is much louder than the noise are almost untouched, but noise reduction always affects the softest parts of a voice, such as breaths and quiet consonants. Compare with the original, and lower the settings if the voice sounds thin or hollow. The loudness step then changes the whole file by one gain, except where the limiter works.

Yes. The WAV keeps the file's sample rate and channels, and both channels get the same noise reduction, so the balance between them does not move. It is 24-bit when the file was 24-bit or more (a 32-bit WAV too), otherwise 16-bit.

Up to 20 minutes of mono or 10 minutes of stereo sound, in a file of up to 50 MB. A 16-bit WAV at 48 kHz reaches 50 MB after about 9 minutes of mono or 4.5 minutes of stereo, so longer recordings have to come as FLAC, MP3, M4A or Ogg. The original, the cleaned version and the removed noise are all kept in memory so that you can compare them: about 1 GB for 10 minutes of stereo at 48 kHz. A phone may not have that much: the page then says that the device ran out of memory, or the browser reloads the tab. Cleaning takes a few seconds for a voice memo and about 20 seconds for 10 minutes of stereo on a current desktop computer; phones are slower. Cut longer recordings into parts with Trim Audio first.

Once the page has loaded, yes. Nothing is fetched while it cleans.

No. The page loads a small script into your browser and every operation runs on your device. You can check this yourself in the browser's network panel: after the page has loaded there are no upload requests. Closing the tab discards the file.

No. These basic tools are free and store nothing on our servers. The only limits are the tool's own: files up to 50 MB; up to 20 minutes of mono or 10 minutes of stereo. Within them, the practical limit is your device's memory.

Nothing you open or type. While the tab is open, the tool pages keep a short list of the steps taken in it, in the tab's session storage: the step (page opened, started, succeeded, first success, used again or failed), the tool's name, the time, how many files were used and, for a failure, an error code; never a file name, the contents of a file or your text. It holds at most the latest 200 steps, and the browser deletes it when the tab closes. The page never sends it. Where a Report a problem button is shown, you can attach the last 20 steps to your report; nothing is attached unless you tick that box. In local storage the page keeps only choices you made: a Light or Dark theme until you choose System again, the language you picked with the EN / 中文 switch if it differs from your browser's language, and this tool in your recently used list if you switched that list on at the tools page.