AI Noise Remover
Clean the background noise out of a voice recording, a podcast, or the sound of a video. An AI model does the work inside your browser, then you compare before and after and choose how much to remove.
Singing mode stops at 75%. Past that, the model starts removing breaths and the quiet end of held notes along with the noise.
Only steady hiss or mains hum? The classic Noise Reducer needs no model download.
How to remove background noise from audio
- Drop in a recording or a video (MP3, WAV, M4A, FLAC, OGG, MP4, MOV, or WEBM).
- Pick Voice for talking or Singing for a sung vocal, then press Remove noise.
- Play the result and switch between Original and Cleaned. Set the strength by ear.
- Download the cleaned audio as WAV or MP3.
How it works
Classic noise tools learn one fixed fingerprint of the noise and subtract it. That works for a steady hiss and fails as soon as the noise changes. This page runs a neural network instead: DeepFilterNet3, an open source speech enhancement model. It listens to the recording 100 times a second and decides, moment by moment, what is voice and what is not, so it can follow noise that comes and goes instead of only a steady hiss.
The model works at 48 kHz, so the top end of a voice is kept instead of being cut off at telephone quality. It runs on your processor, with no graphics card needed. Your browser downloads it once, then your file is decoded, cleaned, and exported on your own device. A stereo file is cleaned one channel at a time, left and right separately.
What it does well
- Spoken voice: podcasts, voiceovers, interviews, voice memos, lessons, video calls, the sound of a talking head video.
- Noise that comes and goes, which a classic noise reducer cannot follow: clicks, bumps, and bursts of background sound.
- Steady noise too: hiss, fans, air conditioning, computer hum.
What it cannot do
- It does not remove echo or reverb. A voice recorded in a bare room will still sound like a bare room, only quieter between words.
- It cannot rebuild a voice that is buried. When the noise is louder than the person talking, expect a thin or patchy result.
- It is built for a voice on its own, not a finished song, and it will damage instruments. To lift a vocal out of a song, use the stem splitter.
- It does not fix clipping, distortion, or a muffled microphone.
- A second person talking in the background is still a voice, so the model may keep them.
Singing needs a light touch
The model was trained on speech. Sung vocals break its assumptions: a held note fades slowly, and a breath before a phrase sounds a lot like noise. When we tested it on sung vocals at full strength, it cut the quiet tails of notes and most of the breaths, and the result measured worse than the noisy original. Limited to 6 dB of reduction, it turned the tails and breaths down by about that much instead of removing them, and the result measured slightly better than the noisy original.
That is why Singing mode starts the strength at 50%, which turns the noise down by at most 6 dB, and stops at 75%. It will not make a noisy vocal silent between lines. It makes it a little cleaner without eating the performance. If you need more than that on a sung take, record it again closer to the microphone if you can.
Getting a natural result
Use the Original and Cleaned switch while the audio plays, and listen to the voice, not the silence. The gaps between words always sound impressive. What matters is whether the words themselves still sound like the person. If the voice turns thin, lispy, or starts to pump, pull the strength down until it stops. A little noise left under a natural voice beats a spotless voice that sounds processed.
If the recording already sounds clean, leave it alone. In our tests on a sung vocal with only a little noise under it, every noise model we tried, this one included, made it measurably worse at full strength, because there was little noise to remove and the only thing left to change was the voice. Clean the noise before you compress or boost the volume, since both make noise louder. And the best fix is still at the source: get closer to the microphone and switch off the fan.
The same job in Xport Studio
Xport Studio, the desktop app we make, has a Noise Remover built in, and it is one of the tools that stay free. It runs on your computer with no browser tab to keep open, and it sits next to the stem splitter, Mix and Master, and the key and BPM finder. The AI voice tools are the paid part: free for 7 days, then $79 once, never a subscription. Download Xport Studio for Mac or Windows.
AI noise remover: common questions
How do I remove background noise from audio?
Drop the recording on this page, pick Voice or Singing, and press Remove noise. An AI model cleans the file on your own device. When it finishes, switch between Original and Cleaned while it plays, set the strength, and download a WAV or MP3.
Is my recording uploaded anywhere?
No. The model and the audio engine are downloaded to your browser once, then your file is decoded, cleaned, and exported on your own device. Nothing is sent to a server, so we never see, store, or train on your recording.
Can it remove background noise from a video?
It cleans the audio track of a video. Drop an MP4, MOV, or WEBM and you get the cleaned sound back as a WAV or MP3, which you can lay back under the picture in any video editor. The page does not export a new video file.
Does it work on singing?
With a light touch. The model was trained on speech, and at full strength it treats breaths and the quiet tail of a held note as noise and removes them. Singing mode starts the strength at 50% and stops at 75% for that reason. If the vocal already sounds clean, leave it alone.
What is the difference between this and the Noise Reducer?
The Noise Reducer is a classic spectral tool: it learns one steady noise, like hiss or a fan, and subtracts it, and it needs no model download. This page uses a neural network that follows noise as it changes, so it copes with sounds that come and go. Use the classic tool for steady hiss and mains hum, and this one for messier rooms.
How long a file can I clean, and how long does it take?
Up to 20 minutes per file. The speed depends on your device, so the page shows the real time it took when it finishes. The first run also downloads the model and audio engine, about 66 MB in total, which your browser keeps for next time.
Who makes this?
We're the small team behind Xport Studio, a private desktop app for AI vocals, stems, and mastering that processes your audio on your own machine. These free tools run on the same idea: your audio stays on your device. If you're curious about the bigger thing, take a look.