Fix distorted audio

Crackly, clipped or crunchy on the loud parts? fix.fm finds the flattened peaks and repairs them. Free, in your browser, nothing uploaded.

What's happening

Distorted audio crackles or crunches on the loud parts. It usually means the recording was too hot: the voice got louder than the mic or recorder could take, so the tops of the sound wave were cut off flat. That's called clipping, and you hear it as a harsh, fuzzy edge on laughs, shouts and hard consonants.

fix.fm looks for runs of samples stuck at the same peak level, which is what a clipped wave looks like up close, and counts how much of the recording is affected. Even 0.05% is enough to hear on a clean voice. Past 1.5%, the distortion is on almost every loud word.

FX-04 / Distorted sample0:10

The distorted sample. The orange bars are where the wave was cut off flat. Those are the crunchy moments you hear.

How far the declipper goes

The declipper fixes one kind of distortion: a wave that was cut off flat because the recorder ran out of room. Distortion that happened before the recorder, in an overloaded preamp or microphone capsule, sounds like a cousin of it and is a different shape, and a file that was clipped and then saved as MP3 is a third. The three cases below are the same clean reading, pushed 9 dB too far in each of those ways, so we know what was there and can measure how much of it came back.

FX-04 / How far the declipper goes7 clips

One clean reading, damaged three ways in software, then run through this page's fixer at full amount. Every clip is matched to -23 LUFS so only the damage differs.

Clean reading

The reference. The loud words stay round and the consonants have a clean edge.

Clipped, 9 dB over

Recorded too hot. The loud words crackle; the quiet ones are untouched.

Clipped, after the fix

The crackle is gone and the file is a little quieter. Listen to the loudest word: it is a rebuilt estimate, not the original.

Overdriven, 9 dB into a saturator

Just as distorted, but warmer and fuzzier, with no crackle. This is what an overloaded preamp or capsule does.

Overdriven, after the fix

Identical to the one above. The check finds no flat tops, so the fixer leaves the file alone.

Clipped, then saved lossy (16 kHz)

The same clipped take with its flat tops smoothed into ripples. It sounds nearly the same as the clipped one.

Clipped and lossy, after the fix

Less crackle than before, more than the clipped fix. The ripples hide exactly where each cut began.

Public-domain LibriVox voice. The clipping is a software hard clip, the overdrive a software saturator and the lossy case a low-pass filter, not a real recorder, preamp or encoder, so the clean version really is what was lost.

Thirty milliseconds of the loudest word, clipped in the recorder: as damaged, after the fix and clean — Pushed 9 dB over full scale and cut off flat. The orange line runs flat along the ceiling wherever the wave was cut off. The black line, showing only where it differs, climbs out past the ceiling and lands on the dashed line, because a cut-off peak was at least as tall as the ceiling and the samples on either side say how much taller. The output itself is held under the ceiling; it is drawn here at the level that matches what was there.

Thirty milliseconds of the loudest word, overdriven before the recorder: as damaged, after the fix and clean — Pushed 9 dB into a soft saturator, so the tops are bent round instead of cut off. The orange line never goes flat. Every peak is bent round on its way up, the quiet parts included, so there is no ceiling to measure and no cut to rebuild from. The black line is hidden under the orange one: the output is the input.

Thirty milliseconds of the loudest word, clipped, then saved as a lossy file: as damaged, after the fix and clean — The clipped take low-passed at 16 kHz, which ripples the flat tops. The flat tops are gone and the orange line ripples above and below where they were. The fixer finds the pile of samples at the old ceiling and rebuilds from a little under it; the black line gets close to the dashed one, less exactly than in the clipped case.

Measured on the fixer's output at full amount
Case	The check	Shape, before to after	Peak after
Clipped in the recorder	1.2% clipped, moderate	15.4 to 26.3 dB (+10.9)	-1.2 dBTP
Overdriven before the recorder	0% clipped, fine	15.5 to 15.5 dB (+0)	0 dBTP
Clipped, then saved as a lossy file	0.9% clipped, moderate	15.3 to 20.3 dB (+5)	-1.2 dBTP

The check: the share of audible samples sitting in cut-off runs, the number this page grades (fine under 0.05%, severe from 1.5%). Shape: how close the wave is to the clean reading, as signal over error in dB with the best level in each 20 ms window, so being quieter is free and a flat top or a wrongly rebuilt peak is not; higher is closer, and the clean reading itself would be infinite. Peak: true peak of the output.

The clipped take is the one the fixer is for. It finds 1.2% of the loud samples cut off, rebuilds them and comes out 10.9 dB closer to the clean reading, graded fine. The overdriven take starts just as far from clean, 15.5 dB against 15.4, and the check reads it as fine, because nothing in it is flat. The fixer passes it through untouched, which is the right call: a declipper that rebuilt peaks that were never cut would be inventing a waveform. For that kind of damage, which also covers a blown speaker or a pedal, nothing on this page helps.

The lossy take is in between. The encoder's smoothing hides where each cut began, so the fixer rebuilds from a little under the old ceiling and gets 5 dB closer rather than 10.9. If you still have the WAV, use it. The clipping vs limiting article goes further into the extreme case where too much of the wave is missing to rebuild, and how we test has the pass marks behind these grades.

How it works here

01
Finds the flat tops
It looks for the loud moments where the wave hit the ceiling and got cut off flat.
02
Redraws the peaks
It rebuilds each flattened peak from the shape of the sound just before and after, then brings the whole file down so the new peaks fit.
03
Stays on your computer
Your file is never uploaded. The repair runs in this browser tab.

Questions

Can distorted audio really be fixed?

Light to medium clipping, yes, often very well. When only the very tops were cut off, there's enough of the wave left to redraw them. When a recording was pushed so hard that most of each word is flat, some crunch will stay.

Is my audio uploaded anywhere?

No. The file is read and repaired inside this browser tab. It never leaves your computer.

Why is the fixed version quieter?

The rebuilt peaks are taller than the ceiling that cut them, so the whole file comes down a little to make room. If it ends up too quiet, pick the "too quiet" fix as well and it will bring the voice back up safely.

Does it work on an MP3 that was clipped?

Yes. Saving as MP3 or AAC smooths the flat tops into small ripples, so fix.fm looks for the pile of loud samples at the old ceiling instead and rebuilds from just under it. If you still have the original WAV, use that: it keeps more of the shape to work from.

Does it work on distorted music or a blown speaker sound?

It's built for clipping, the flat tops you get from recording too loud. It's tuned on voices. Distortion from a damaged speaker, a cheap mic or an effect pedal is a different kind of damage and won't be undone.

How do I stop audio clipping next time?

Turn the input gain down until your loudest laugh peaks around -6 dB on the meter. A little quiet is easy to fix. Clipped is much harder.

OFTEN TURNS UP WITH DISTORTED