Remove background noise from any audio

Q: Does it work on music?

CleanGrab is tuned for voice and speech — podcasts, interviews, calls, voice memos, lectures. On music or other non-speech audio the result can be poor, and the tool will tell you when it couldn't detect clear speech instead of handing you a mangled file.

By ShinobiTools Team · Last updated: July 2026

Drop a noisy voice recording in and get a clean version back — hum, hiss, fans, traffic and keyboard noise removed on our own hardware. Free, no account, no limits.

Clean a recording, free →

Built for voice, not a generic filter

CleanGrab runs DeepFilterNet3, an open-source speech-enhancement model, on our own hardware, so there's no per-file fee and no watermark: drop in a recording and get a cleaned voice back as a 192 kbps MP3 plus a lossless WAV. No account, no sign-up, and your upload is deleted the moment it's processed.

Hum and electrical buzz
Hiss and background static
Fans and air conditioning
Traffic and street noise
Keyboard clatter and typing

How to remove background noise

Open CleanGrab and drop your recording in (MP3, WAV, M4A, FLAC, OGG, or a video file).
Pick a strength — Gentle, Balanced or Strong.
DeepFilterNet3 separates your voice from the noise on our own hardware.
Compare before/after in your browser, then download the cleaned MP3 or WAV.

What it won't fix

Heavy room echo and reverb are a harder, separate problem: CleanGrab reduces light echo, but it will not fully erase a very reverberant room. We'd rather say that plainly than promise something the tool can't do.

CleanGrab is tuned for voice and speech — podcasts, interviews, calls, voice memos, lectures. On music or other non-speech audio the result can be poor, and the tool will tell you when it couldn't detect clear speech instead of handing you a mangled file.

What the three strengths actually change

Gentle, Balanced and Strong aren't vague "more or less" settings — each one is a hard ceiling on how much the noise may be attenuated, measured in decibels:

Gentle caps the reduction at about 6 dB. The noise floor drops audibly but stays present, and your voice keeps all of its body. This is the setting for a recording that's already decent and just a bit hissy.
Balanced caps it at about 12 dB, which is the default because it's the point where most recordings sound clean without sounding processed.
Strong removes the cap entirely and lets the model take out as much as it thinks it can. It's the right answer for a genuinely noisy recording — a fan running, traffic through an open window — and it's also where artefacts appear if the noise wasn't that bad to begin with.

So the practical recipe is: start on Balanced. If noise is still distracting between words, go Strong. If the voice sounds thin, hollow or slightly underwater, drop to Gentle — that sound is the model removing parts of your voice along with the noise, and less processing is the fix, not more.

One thing to listen to specifically: the tail of an "s" or a "t". Sibilance sits in the same frequency range as hiss, so over-processing dulls those consonants first. If the words start sounding lisped, you've gone one step too far.

Noise the model was built for, and noise it wasn't

DeepFilterNet3 was trained to pull a speaking voice out of added background sound. That's why it's excellent at some things and structurally unable to do others.

Removed well, because it's steady and unlike a voice: fans and air conditioning, computer and fridge hum, mains hum, tape and preamp hiss, traffic and street noise, rain, keyboard and mouse clatter, distant machinery.

Not removed, and worth knowing why:

Another person talking. That's speech, so the model protects it exactly as carefully as it protects you. Two voices at once is a separation problem, not a noise problem.
Music under speech. Same reason — it's structured, tonal audio, and pulling it out without shredding the voice is a different task.
Heavy reverb. The echo is your own voice arriving late. It's reduced, not erased, and a very live room can't be undone.
Clipping and distortion. If the input was recorded too loud, the waveform was flattened at the moment of recording and the information is gone. No model rebuilds it — it can only make the distortion cleaner-sounding.
Wind on a bare microphone. Wind overloads the capsule itself, producing a low rumbling that isn't really "noise added to" your voice. A foam windscreen at record time is the only real fix.

If your recording is mostly music or has no clear speech in it, you get told so rather than handed a mangled file — a wrong tool honestly reported beats a bad result silently delivered.

Common questions

What kinds of background noise does it remove?

Hum, hiss, fans, air conditioning, traffic, street noise and keyboard clatter are removed well, while the voice is kept natural.

Can it fix echo or reverb?

Does it work on music?

Is it really free and unlimited?

Yes — no account, no watermark, no daily or monthly cap. Adobe Podcast, Krisp and Cleanvoice all limit free usage to push you toward a paid tier; CleanGrab has no paid tier and no time limit. The ads around the tool cover running our hardware.

What happens to my file?

No. Your upload is deleted the moment it's processed and the cleaned file is wiped from the server within 45 minutes. Nothing is stored, shared or used to train anything — see privacy.