Remove background noise from audio
To remove background noise, drop a recording here and press Remove noise: an open-source AI model cleans hum, hiss, fans and street noise out of the voice right in your browser. Listen to before and after, then download MP3 or WAV. Free, no signup, nothing is uploaded.
Record from your microphone, then clean the take. The browser's own noise suppression is switched off here, so you hear the room as it really is first.
| Before | After | |
|---|---|---|
| Background noise | - | - |
| Voice above the noise | - | - |
The first run downloads the noise model, about 12 MB. Your browser keeps it for next time.
Removes steady noise: hum, hiss, fans, traffic, rain, keyboards. Music, other voices and room echo mostly stay.
Nothing is uploaded. The recording is cleaned in this tab and gone when you close it.
How the noise remover works
The cleaning is done by DeepFilterNet 3, an open-source speech enhancement model. The page downloads it once, runs it in a background thread of your browser and feeds it the recording 10 milliseconds at a time. For every slice it works out which parts of the sound are voice and which are noise, and turns the noise down as far as the strength allows: Light stops at 6 dB, Medium at 12 dB, Strong at 16 dB. There is no stronger setting on purpose: beyond that, on a hum or loud noise, the model starts cutting holes in the voice. The cleaned voice is then brought back to the level it had, so before and after compare fairly.
Two tips. A recording made close to the microphone cleans far better than one made across a room, because there is more voice to keep. And start with Medium or Light: the less noise a model removes, the less it can bend the voice. If all you need is a transcript, you can skip cleaning: Speecho transcribes noisy recordings as they are, and the audio quality checker tells you whether yours needs help at all.
Questions
Does my recording leave my device?
No. The file is read, cleaned and turned into MP3 inside this browser tab. The only download is the noise model itself, from speecho.app. You can watch the network tab of your browser: the recording is never sent anywhere.
What noise does it remove, and what stays?
It is made for steady and repetitive background sound: air conditioning, fans, computer hum, the hiss of a cheap microphone, traffic, rain, keyboard clicks. Music, a television, other people talking and the echo of a large room mostly stay, because to the model they sound like voice too.
Which strength should I use?
Medium suits most recordings. Light keeps a little of the room, which sounds more natural in interviews and on location. Strong is for a loud hum or hiss. Even Strong leaves a little noise in the pauses on purpose: removing all of it makes the voice break up. You can change the strength and run it again; the original file is never changed.
Why is the cleaned file mono, at 48 kHz?
The model works on one channel at 48 kHz, so stereo is mixed down and other sample rates are converted first. For a voice that loses nothing you would hear. The MP3 is 128 kbps; the WAV is 16-bit, for editing.
How long can the recording be?
Up to 60 minutes per file, and up to 30 minutes when you record on the page. The whole recording is held in the tab's memory, so on a phone or tablet the limit is 15 minutes for both. A recent laptop cleans an hour in a few minutes; a phone is several times slower.
How is this different from online AI enhancers?
Most online enhancers, Adobe Podcast Enhance among them, upload your file to their servers and ask you to sign in, and many rebuild the voice with generative AI, which can change how it sounds. This tool only turns the noise down, keeps the voice as it was recorded, and never uploads the file.
Should I remove noise before transcribing?
Usually not. Speech recognition copes well with steady noise, and heavy noise removal can thin out the consonants it relies on. Transcribe a sample first and clean only if words go missing; this guide to fixing bad audio shows which problems are worth fixing.