Narration Clean-up: remove mouth clicks, breaths, sibilance

Where it runs

Works in a current Chrome, Edge, Firefox or Safari, on Windows, Mac, Linux and Chromebook computers, Android phones, iPhones and iPads. Nothing to install, no account.

The first file loads the audio engine (ffmpeg, about 31 MB) once, and your browser keeps it. Very long books are easiest on a computer, which has more memory than a phone.

4 simple steps.

Drop your narration

One take or a whole book: WAV, FLAC, MP3 or M4A, straight from your editor. Each file is listened to on your device.

See what was found

Mouth clicks, breaths, harsh “s” sounds, plosive pops and mains hum, counted and marked on a timeline of the recording.

Hear it before and after

Twenty seconds re-render as you change a setting. Switch between Original and Cleaned while it plays, or jump to the next breath or click.

Clean and download

Only the moments it found are touched. Save as WAV, FLAC or MP3, then master for ACX.

What it fixes

ACX asks for audiobook files “free of extra sounds like plosives, microphone pops, mouse clicks, excessive mouth noise, and outtakes”, because they “distract listeners and can lead to bad reviews”, and for a sound that stays consistent from file to file.¹ Those are exactly the things a level meter can’t see. This tool finds five of them:

Problem	What you hear	How it’s fixed
Mouth clicks	Ticks, smacks and wet clicks from lips and tongue, worst in pauses and at the start of phrases	The band above 1.5 kHz is lowered 24 dB for the click’s few milliseconds
Breaths	Loud inhales between phrases	The breath is lowered as a whole (−6, −12 dB, or to the room tone), with 25 ms fades
Sibilance	Harsh, hissing “s”, “sh”, “z”, “ch”	A dynamic EQ at your own sibilance frequency, active only during the hiss
Plosive pops	A low thump on “p” and “b”	The band below 150 Hz is lowered 18 dB for the pop alone
Mains hum	A steady 50 or 60 Hz drone, sometimes with a buzz of harmonics	6 Hz-wide notches at the hum and any strong harmonics

Outtakes, re-reads and long gaps are an editing job, and background noise, loudness and peaks are mastering’s: run ACX mastering after clean-up.

How each fix works

Mouth clicks

A mouth click is a burst of a millisecond or two, mostly above 2 kHz. The detector looks at the sound above 2.5 kHz in half-millisecond steps and marks a burst when it is short (4 ms or less), stands out from the 15 ms around it (by 15 dB on Normal), and is louder than anything 3 to 15 ms either side. That last test matters: in voiced speech, every vocal-fold pulse is a small burst too, but pulses repeat every few milliseconds while a click stands alone. Gentle asks for 18 dB, Strong for 12.

FFmpeg has a de-clicker, adeclick, made for scratches on records.² In our tests it removed about 1 dB of a mouth click, and at more sensitive settings it began to alter the voice, so it isn’t used. Instead, the recording is split into bands with FFmpeg’s crossover filter, whose bands add back up flat, and only the top band is turned down for each click, with 1.5 ms fades. A detected click drops by about 16 dB; on Normal, under 1% of the recording is touched at all.

Breaths

Breaths are found in 10 ms frames: not voiced (no pitch), quieter than speech by at least 10 dB, louder than the room tone, 120 to 900 ms long, with a pause before or after, and not a sibilant (most of their energy is below 4 kHz). The whole breath is lowered, with 25 ms fades so it never pumps. However far you set it, a breath is never pushed below the room tone: a hole of digital silence in the middle of a pause sounds worse than the breath.

Sibilance

Everyone’s sibilance sits at a different pitch. The analysis tries four detection frequencies (4, 5, 6 and 7 kHz) and keeps the one that best tells your “s” sounds from your vowels. The threshold is then set from your recording: just under your loud sibilants, and never below the level your vowels reach there, so ordinary speech is left alone. The cut is a bell around that frequency (FFmpeg’s adynamicequalizer), with a 1 ms attack and 60 ms release.² On our test recordings, Normal lowered “s” sounds by 4 to 6 dB (the loudest quarter by 7 to 11 dB) and changed other speech by 0.0 dB; Gentle and Strong move about 2 to 3 dB either way.

Plosive pops

A pop is a burst below 120 Hz that is at least 14 dB louder than that band usually is in your speech, carries most of the sound at that moment, and lasts 10 to 150 ms. Only the band below 150 Hz is lowered (18 dB, with 8 ms fades), from just before the pop to just after it.

Mains hum

Hum shows up as a needle-thin peak at 50 or 60 Hz, depending on where you live, often with harmonics above it. The analysis looks for it in the quietest half-seconds of the recording, where nothing covers it, and in the recording as a whole. Only harmonics within 20 dB of the strongest are notched: weaker ones are inaudible, and MP3 files, which drop quiet frequencies, can make tiny ones look like peaks. Each notch is 6 Hz wide, narrow enough that the voice around it doesn’t change.

How it was tested

Every number above was measured, not estimated. The test material was real narration: chapter 2 of F. Scott Fitzgerald’s The Curious Case of Benjamin Button, read by Mike Vendetti, and “The Sisters” from James Joyce’s Dubliners, read by a LibriVox volunteer, both public-domain recordings.³ We added clicks and pops at known places, ran the files through the same audio engine your browser uses, and compared the result with the original sample by sample.

Fix (Normal)	Result	Everything else
Mouth clicks	Found: about 9 in 10 clicks that stood out 12 dB or more (87–91%); lowered about 16 dB	0.4–1.7% of samples touched, by strength
Breaths (−12 dB)	286 breaths in a 20-minute chapter, lowered by a median 12.1 dB	Speech between them unchanged
Sibilance	“s” sounds −4 to −6 dB (median)	Other speech 0.0 dB
Plosive pops	50 of 50 added pops found; lowered 14–15 dB	Only the band below 150 Hz, only at the pops
60 Hz hum	−33.5 dB	100 Hz–3 kHz unchanged (0.00 dB)

In Chrome on our test machine, a 20-minute MP3 chapter took about 30 seconds to analyse and 30 seconds to clean; your browser and device set the pace.

Where it fits

Record, ideally with a pop filter, a little off-axis, with water at hand (dry mouths click more).
Edit: take out mistakes and re-reads, and tidy the pauses.
Clean up here. Keep WAV or FLAC so nothing is lost before mastering.
Master with ACX mastering: loudness, peaks, noise floor and room tone, out as 192 kbps MP3.
Check with the ACX checker, and listen through.

Editing in Audacity? Audacity for audiobooks shows where this fits next to its own effects. Recording at home? The home recording guide covers the room and the microphone, and the noise floor meter measures how quiet your space is.

What it can’t do

Clicks inside loud vowels. They are mostly masked by the voice, and telling them from it is unreliable, so they are left alone.
Room echo, traffic, computer fans. Noise under the voice needs a quieter room; mastering can only turn the pauses down.
Clipping and distortion. Re-record the line.
Judgement. A breath that belongs to the performance, a lip smack that is part of a character: only your ears know. Listen to the preview, try Gentle, and turn off what you don’t want.

Sources

What are the ACX audio submission requirements?, ACX Help, checked 23 October 2026
FFmpeg filters documentation: adynamicequalizer, acrossover, bandreject and adeclick
LibriVox recordings on the Internet Archive: The Curious Case of Benjamin Button and Dubliners (public domain)

Questions, answered.

How do I remove mouth clicks from a recording?

Drop the file here and leave Mouth clicks on Normal. Each click is found by its shape (a burst of a few milliseconds that stands out from the sound around it and from its neighbours) and only the band above 1.5 kHz is lowered for those milliseconds, so the click drops by about 16 dB and the voice around it is untouched. Gentle finds fewer, Strong more.

Should I remove breaths from an audiobook?

Lower them rather than delete them. Narration with every breath cut out sounds unnatural, and cutting them changes the timing. The usual approach is to make breaths quieter so they stop drawing attention: −6 or −12 dB here, or down to your room tone. The tool never lowers a breath below the room tone, so a pause never turns into dead silence.

What is a de-esser?

A de-esser turns down sibilance, the hiss of “s”, “sh”, “z” and “ch” sounds, only while it happens. This one is a dynamic EQ: it listens at your voice’s own sibilance frequency (found from your recording, between 4 and 7 kHz) and cuts that region when an “s” is louder than your vowels ever are there. On Normal, “s” sounds come down about 4 to 6 dB and other speech is unchanged.

What are plosives, and can they be fixed after recording?

Plosives are the puffs of air from “p” and “b” that hit the microphone as a low thump. A pop filter and speaking slightly across the microphone prevent them. Afterwards, the thump can be lowered: it lives below about 150 Hz, where the voice has little, so the tool lowers only that band for the pop’s 10 to 150 milliseconds.

Will this make my audiobook pass ACX?

It deals with what ACX calls extra sounds, plosives, microphone pops and excessive mouth noise, which a meter can’t check. ACX also needs loudness between −23 and −18 dB RMS, peaks under −3 dB and a noise floor under −60 dB: clean up first, then run ACX mastering, which sets those and measures the result.

Does it change the loudness or remove background noise?

No. Loudness, peaks and background noise are mastering’s job, and the ACX mastering tool does them. Clean-up only touches the clicks, breaths, sibilants, pops and hum it finds; the rest of the recording comes out exactly as it went in, apart from a tiny phase shift where the bands are split.

Should I clean up before or after mastering?

Before. Mastering raises the level of everything, mouth noise included, and its limiter reacts to pops. Clean the edited takes (WAV or FLAC keeps every detail), then master them.

Are my recordings uploaded?

No. Everything runs in your browser on your device: the audio engine downloads once, and your files never leave your computer.

Does it work for podcasts and voice-overs?

Yes. Anything spoken into a microphone has the same problems. The detectors were tuned on audiobook narration, so a single voice works best; music under the voice can confuse them.