How to Remove Background Noise From Audio: RNNoise, Recording Tips and Honest Limits

A podcast recorded next to a humming refrigerator, a lecture under an air conditioner, a voice memo with traffic behind it — the words are all there, but the noise sits in every pause and makes the whole recording feel amateur. Our free AI audio noise remover strips fan hum, hiss and background chatter from voice recordings, free and unlimited, with the AI running entirely on your device: your audio never leaves the page. This guide explains how to use it, how the RNNoise model actually works, how to record so the cleanup is easy, and the honest limits — what it will not fix, no matter how good the model is.

How to Use the Audio Noise Remover

Step one: upload your recording. Drop an audio file onto the tool or click to browse — MP3, WAV, M4A or OGG all work, and voice recordings work best. The file stays on your device; nothing is uploaded. Step two: let the AI load, then remove the noise. The RNNoise model loads once, then denoises any recording in seconds. Step three: compare and download. Use the before and after players to A/B the result honestly, then download a clean WAV. There is no account, no watermark and no file limit — run as many recordings through it as you like. One practical note: the model loads once per session, so the first file takes a little longer while the roughly 1.5 MB WebAssembly build arrives; after that, processing is fast, even in real time on a phone.

How RNNoise Actually Works

RNNoise is a small recurrent neural network trained on thousands of hours of speech mixed with real-world noise, and it is the same open-source technology behind many professional denoisers. It processes audio in short frames of about 10 milliseconds, estimating for each frequency band how much of the energy is voice and how much is noise, then suppressing the noisy bands. Because it is a speech model, it preserves formants and consonants — the parts that make words intelligible — while pushing steady background sounds down 15 to 25 dB. The whole network compiles to roughly 1.5 MB of WebAssembly, small enough to run in real time on a phone, which is why this tool can run entirely in your browser with no server at all. That on-device property is not a gimmick: a voice recording is personal, and uploading it to a stranger's server just to remove a fan hum would be a strange trade.

Worked Example: The 42-Minute Podcast by the Fridge

The tool page's own example is a 42-minute podcast recorded next to a humming refrigerator. In the original, voice peaks hit −12 dB while the fridge hum sits at −38 dB — clearly audible in every pause. After processing, the hum drops to about −58 dB, a 20 dB reduction that makes it effectively inaudible, while voice clarity is unchanged. The numbers on the files are worth knowing before you start: the 8.4 MB MP3 goes in, and a 440 MB WAV comes out, because WAV is uncompressed — one minute of stereo 44.1 kHz audio is about 10 MB. For publishing, compress the result to 128 kbps MP3, which brings that 42-minute episode to about 40 MB. So the workflow is: denoise to WAV for maximum quality, edit if you need to, then compress once for sharing. Compressing first and denoising second would bake MP3 artifacts into the recording that no denoiser can remove.

Recording Tips for the Cleanest Result

Denoising works best on a decent recording, so five habits matter more than any post-processing. First, record in the quietest room available — a closet full of clothes beats a big empty room that sounds treated but is not. Second, keep the mic 4 to 6 inches from your mouth; every doubling of distance roughly halves the voice-to-noise ratio, so distance is the single biggest lever you have. Third, use a directional mic if you have one, because it rejects sound from behind and beside you before the model ever sees it. Fourth, record WAV at 44.1 kHz or higher — heavy MP3 compression bakes in artifacts no denoiser can fix. Fifth, leave two seconds of room tone at the start of every recording, so you can A/B the before and after honestly and hear exactly what the model removed. Follow those five, and the model has an easy job; ignore them, and you are asking software to rescue a recording that was never really captured.

What It Removes Best — and What It Will Not Fix

Constant sounds are the model's home ground: fan hum, air-conditioner drone, electrical hiss, keyboard clatter and distant traffic all fall away cleanly, because steady noise is easy to estimate band by band. Intermittent sounds are a different story: a door slam, a dog bark, or someone talking over you are only partially reduced by any denoiser, because they look too much like speech for a short window. Heavy room echo stays as well — the model removes additive noise, not reflections, so if your recording sounds echoey, re-record closer to the mic in a soft-furnished room, because no post-processing fully fixes reverb. And music is out of scope entirely: RNNoise is trained for speech, so singing, instruments and sound effects get distorted. Keep it to voice recordings — podcasts, voiceovers, meetings, lectures and interviews — and it will do that one job very well.

Common Mistakes

Uploading a recording to a random online denoiser. Your voice recording is personal; this tool exists so it never has to leave your device. Feeding it music. A speech model distorts music — use it for voice only. Expecting echo removal. Echo is reflection, not added noise, and it stays; re-record closer in a softer room instead. Compressing before denoising. Heavy MP3 compression bakes in artifacts; denoise the best-quality file you have, then compress once for publishing. Recording far from the mic. Every doubling of distance halves the voice-to-noise ratio, and no model fully recovers that. Panicking at the WAV file size. Uncompressed WAV is large by design — about 10 MB per stereo minute — and that is maximum quality for editing; compress to MP3 or AAC afterwards for sharing. Processing very long files on a phone. Files over about 100 MB may strain mobile browsers, and recordings over an hour are more comfortable on a desktop browser.

Limits Worth Knowing

The honest limits are four. The model is speech-only, so music, singing and sound effects distort. Intermittent loud sounds are only partially reduced. Room echo and reverb are not removed. And output is uncompressed WAV at the recording's original sample rate, so long recordings produce large downloads — the 42-minute example yields a 440 MB WAV — and files over about 100 MB may strain mobile browsers. Within those limits, the tool does one job completely: it takes a voice recording with steady background noise and returns the same voice, cleaner, without an account, a watermark, a file limit or an upload. The companion guides are our voice recording guide, for the mic setup, trimming and WAV download that produce a clean source file, our subtitle editing guide, for captioning the cleaned recording, and our YouTube thumbnail guide, for the image that gets the finished video clicked.

Frequently Asked Questions

More guides from ToolNest