How to Remove Mouth Clicks from Audio Without Damaging Speech

To remove mouth clicks from audio without damaging speech, repair each event according to where it lands, not according to how annoying it sounds. A short click in the gap between words can often be repaired locally, using a tiny bridge for an impulse or matching room tone for a longer click. The same edit across a t, d or k can punch a hole in the word. That second case needs a processor tuned for mouth noise or a careful time-frequency repair.

I built a controlled test to measure the difference. Seven copies of one real CC0 teeth click were placed over the same LibriVox speech recording: four in measured low-energy gaps and three directly over spoken audio. A targeted repair removed the four gap clicks while leaving every sample outside those selected intervals untouched. A generic whole-file de-click pass barely moved this broader mouth transient and changed more of the surrounding file.

The useful lesson isn't that one algorithm lost. It's that “mouth noise” is not a preset. Diagnose the transient, expose the removed signal, and stop before wanted articulation enters it.

The Short Answer: Remove the Smallest Defect You Can Isolate

• Click in a pause: zoom in, select only the transient and use a tiny interpolation, redraw, replace or silence edit. Audition both boundaries.
• Click across speech: use a specialist Mouth De-click processor or spectral repair. Do not mute the interval.
• One click inside a paid line while the speaker is available: record a pickup before forcing an aggressive repair. A clean re-take is often faster and safer than rebuilding a vowel.
• Clicks throughout a clean voice track: test a conservative whole-track pass on a duplicate. If available, audition the residual or “clicks only” output for missing consonants; also compare the repaired speech with the source.
• Needle-thin click at an edit or digital dropout: follow the current Sound Forge clicks and pops guide; that is a different fault family.
• Low thump on a p or b: use plosive repair or a local low-frequency edit, not mouth de-click.
• Long bright s sound: use de-essing. A de-clicker is looking for a short event, not sustained high-frequency energy.

If the recording also has hum, room reflections, clipping or broadband noise, start with the audio restoration diagnosis hub. It keeps you from stacking unrelated processors on a problem you haven't even named yet.

How to Tell a Mouth Click from the Other Sharp Noises

A mouth click is a brief oral transient produced around tongue, teeth, lips and saliva movement. It may be a single spike, a small asymmetric burst or a sticky cluster with a short decay. Unlike a digital click caused by an abrupt discontinuity, it doesn't have to be one sample wide. It also isn't a plosive's broad low-frequency air blast, and it isn't sibilance either: it doesn't stretch across the whole s or sh.

Listen at normal speed first, then zoom far enough to see the event without losing the surrounding word. A waveform identifies the timing and amplitude jump; a spectrogram shows whether the event cuts vertically across a broad band, blooms mainly at low frequencies or continues as high-frequency energy.

Use these checks before touching a control:

1. Does the sound occur before, after or inside a word? Timing determines whether a local edit can be invisible.
2. Is it one event or dense crackle? A cluster may need click widening, de-crackle or several local repairs rather than one interpolated line.
3. Does the waveform jump at an edit boundary? Move or crossfade the edit before blaming the speaker.
4. Does it vanish when you change the playback device? Then the file may be clean. Check the playback path with the distorted-audio fault test.

Don't diagnose from a boosted top end alone. Compression, limiting and bright EQ make small clicks easier to hear, but they say nothing about what caused them. Repair first; master later.

A Controlled Test: Seven Real Clicks over One Voice

The speech is an 11.12-second excerpt from Claudia Caldi’s LibriVox reading of “Epitaph on a Sailor” in Short Story Collection Vol. 106. The contaminant is the 12 ms core of colorsCrimsonTears’ CC0 “Click” recording, made by clamping teeth together. LibriVox labels its recordings public domain in the United States and advises listeners elsewhere to verify local copyright status. The inserted click stands in as a controlled proxy for a short oral transient; nobody is claiming every lip smack or saliva noise has the same shape.

I inserted the click seven times at −14 dBFS. Four events sit in low-energy gaps; three overlap high-energy speech. The local method replaces only a 16 ms selected interval around each gap click with a straight bridge between its edges. The three over-speech clicks are deliberately left alone because deleting their interval would delete voice too.

For the generic baseline I ran the FFmpeg 8.1.2 adeclick filter in overlap-save mode at thresholds 4, 3 and 2. The current filter documentation covers the same window, overlap, autoregression, threshold, burst and overlap-method controls used in the test. It's an open, reproducible impulse-removal baseline; it doesn't stand in for a processor designed specifically for mouth noise.

The residual-error levels and outside-window error are relative to clean-speech RMS, where lower is better. Click-window improvement and SI-SDR run the other way: higher is better. STOI is a separate intelligibility estimate. Overall SI-SDR improved from 17.2373 dB to 20.9168 dB after the gap repairs even though the three clicks over words remained. STOI stayed at 0.991559. The aggressive generic pass reached 17.3590 dB SI-SDR and 0.990857 STOI. These measurements compare each file with the known clean reference; listening to consonants, room tone and repair boundaries still gets the final word.

The outside-window figure needs careful reading. −35.34 dB doesn't mean the output was “35 dB worse.” It means the generic processor made detectable changes outside the known click windows at an error level 35.34 dB below the clean speech RMS, while delivering only 0.189 dB of click-window improvement. That's a poor risk-to-reward result for this source, though it says nothing about every de-clicker or every mouth click.

Hear the Original, Clicks and Two Repairs

Use headphones and keep the playback level fixed. The four MP3 files are listening previews, not the files used for scoring. The test uses 48 kHz mono audio, with 24-bit WAV outputs and the measurement boundary explained in the reproducibility record above. Listen once for the seven clicks, then again for word endings and consonant attacks.

Prevent Mouth Clicks Before Recording

The U.S. National Institute of Dental and Craniofacial Research recommends water as one way to ease dry mouth, but that's health guidance; it doesn't prove one sip will remove a recorded click. If persistent dry mouth affects speaking or swallowing, its guidance says to ask a dentist or doctor about the cause. For a session, keep ordinary water available and record a short monitored test after the speaker has settled.

RØDE’s current mouth-noise guidance recommends testing an off-axis microphone angle and notes that sensitive studio microphones can emphasize mouth noise. Treat its 45-degree example as a starting point instead of a mandatory position: compare a small angle and a little more distance through the actual recording chain, then keep the version that preserves clarity.

A pop filter addresses the plosive air bursts described in the RØDE guidance, not clicks already generated inside the mouth. If a sharp sound has a room tail, listen for reflections too: treating the room may reduce the repeated sound, but it will not prevent the original oral click. Check the result with a short recording.

For voice work, save the test take and its mic position with the session notes. The Sound Forge voice-over workflow covers recording, monitoring and delivery, while the podcast workflow handles repeated spoken-word production.

How to Repair a Mouth Click in a Pause

1. Duplicate the file or save a new version. Do not make the only copy your experiment.
2. Mark the click at normal speed. If you hunt only at extreme zoom, it is easy to repair a harmless waveform detail and miss the audible event.
3. Zoom to the shortest complete transient. Keep the selection inside the low-energy gap and include no more decay than necessary.
4. Choose the smallest suitable repair. Interpolate or redraw a very short impulse; replace or silence only when the surrounding gap makes that edit natural.
5. Audition before, through and after the boundary. A repaired centre can be quiet while either edge clicks.
6. Return to normal zoom. The edit has to sound invisible in the sentence, not merely look smooth under a microscope.

In Sound Forge 2026, the native command is Tools → Repair → Interpolate. The official Interpolate documentation says the selection should be less than 2 ms in most cases; with no selection, the cursor is treated as the middle of a glitch shorter than 1 ms. That's the right scale for a tiny impulse, and it's not permission to interpolate a 12 ms mouth sound across a syllable. Use the full native repair walkthrough for Interpolate, Pencil and DeClicker/DeCrackler controls.

For a longer click entirely inside a pause, use matching room tone instead of stretching a tiny-interpolation tool beyond its purpose. Copy a clean gap from the same recording, replace an equal-length span without shifting later speech, and blend only the patch edges if needed. Keep both transitions inside the pause and listen through the complete sentence. Do not paste digital silence into an audible room bed or use this replacement across a word.

Audacity’s official repair guide makes the same scale distinction in another editor: automatic Click Removal expects a selection of more than 4,096 samples, while Repair is for an individual selection no longer than 128 samples. The numbers are tool-specific, but the principle travels: use a local repair for a local defect.

How to Remove a Mouth Click over Speech

When the click shares time with a consonant or vowel, silence and straight-line replacement remove wanted speech along with the defect. Start with a processor designed for mouth noise. iZotope currently lists Mouth De-click in RX 12 Standard and Advanced, not Elements, and describes it as tuned for clicks and lip smacks on long selections or individual events. For settings that keep speech natural, see the RX 12 Mouth De-click guide.

That product distinction matters. iZotope’s legacy De-click plug-in help separates Click, Thump and Discontinuity types; those labels are not a promise about your current Mouth De-click interface. It also warns that excessive sensitivity can damage wanted audio. Its Clicks only output exposes the difference between original and processed audio. If speech appears in that residual, back off.

A safe specialist pass looks like this:

1. select one difficult phrase plus a clean phrase;
2. start at low sensitivity and no extra click widening;
3. preview at matched level;
4. listen to the removed signal or delta where available;
5. raise sensitivity only until the target click is caught;
6. stop if t, d, k, breaths or vocal texture enter the residual;
7. switch off Listen, Clicks only or delta output, confirm that normal repaired speech is playing, then process the smallest sensible region before considering the whole track.

iZotope’s guide to digital clicks and mouth noise recommends the dedicated Mouth De-click tool for oral noise and local use when events are sporadic. Its vocal-cleanup guide likewise treats stubborn mouth clicks as targeted work in the standalone editor rather than a reason to flatten an entire vocal.

Another candidate is Greenlit Audio’s Mouth De-Click, with free processing and paid controls. Its current setup notes limit processing to 44.1, 48 and 88.2 kHz; other rates pass through dry. It reports 400 ms latency. Audacity offline application can leave the first couple of seconds untreated, and versions before 0.3.1 could return the whole selection unchanged. Test the actual render and installed version, not just the repair counter. This plug-in was not part of the benchmark above.

If Mouth De-click still leaves one click, use spectral repair on that event. A mouth click often appears as a short vertical feature, but select only the energy that does not belong to the surrounding consonant. Process once, compare, and undo if the word loses definition.

Repair Mouth Clicks in Sound Forge

Boris FX positions Sound Forge as a dedicated environment for detailed audio work beyond the timeline. In this workflow it handles close waveform inspection, markers, native repair, plug-in hosting and final assembly around one file, even when a specialist repair runs in another application.

1. Open a copy of the original WAV and create markers for representative clicks.
2. Repair the obvious gap clicks with the smallest native edit.
3. For clicks over speech, send a duplicate selection to RX standalone, or test a compatible installed plug-in if it appears in your Sound Forge setup.
4. Before rendering, switch off any removed-signal monitoring mode. Reimport normal repaired speech without changing sample rate, channel count or start time.
5. Level-match bypass and processed playback.
6. Check every marked consonant and every edit boundary on headphones and monitors.
7. Export a new lossless master; do not overwrite the capture.

Don't assume an RX license, plug-in format or menu path is present just because another user has it. RX 12 Elements includes generic De-click; the current edition comparison places specialist Mouth De-click in Standard and Advanced. If your host does not recognize the plug-in and your RX licence includes the standalone editor, use that editor for the round trip. RX Elements does not include the standalone editor; check the edition before choosing this route.

Run noise reduction after mouth-click repair when practical. Broadband denoise can blur the transient without removing it; compression and limiting can then make the remainder more obvious. For the next stage, use the background-noise guide or compare conventional repair with the tested AI audio restoration workflow.

Five Repairs That Usually Make the Voice Worse

• Silencing inside a word. The click disappears because part of the word disappears.
• Turning sensitivity up until the track looks clean. Visual cleanliness can hide softened consonants and dulled attacks.
• Running several blind passes. Each pass sees the artifacts created by the previous one. Repeated processing can trade clicks for lisping or dullness.
• Using de-essing for every bright noise. De-essing controls sustained sibilant energy; it is not a precision mouth-click repair.
• Compressing before cleanup. Compression followed by makeup gain can bring quiet clicks forward, and dynamics processing may change their shape.

Perfectly sterile dialogue isn't always the best dialogue. Remove distractions and obvious client-delivery defects, but keep natural breaths, articulation and presence. A casual podcast doesn't need the same cleanup density as an audiobook pickup or an exposed commercial voice-over. If the speaker is still available and one important word needs heavy reconstruction, record the pickup.

Mouth-Click Removal FAQ

What causes mouth clicks in audio?

Mouth clicks are short oral transients associated with movement of the tongue, teeth, lips and saliva during speech. Close, bright microphone placement and later compression or high-frequency boost can make them more obvious, but they are different from digital dropouts and edit-boundary clicks.

Can Audacity remove mouth clicks?

Audacity can repair some clicks. Its official guide uses Click Removal for a larger selection and Repair for an individual selection up to 128 samples. A click over speech may still need a specialist mouth-noise processor or careful spectral repair.

Can Sound Forge remove mouth clicks?

Sound Forge can repair tiny isolated defects with Interpolate or Pencil and can run its current DeClicker/DeCrackler tools. For mouth clicks embedded in speech, test a specialist Mouth De-click processor in a compatible plug-in setup or standalone editor, then reimport the result.

Should I remove every mouth noise?

No. Remove distracting clicks and delivery failures, but keep natural articulation and breath texture. Processing every small oral sound can make speech dull, lisping or unnaturally sterile.

Does a pop filter stop mouth clicks?

A pop filter mainly reduces air blasts from plosive consonants. It does not remove clicks generated inside the mouth. Water may ease dry mouth, while mic angle, distance and a short monitored test address the recording path.

Should mouth de-click run before noise reduction?

Usually repair obvious clicks first. Broadband denoise may smear their shape, while compression or limiting can make them louder. Test the order on a short duplicate when the recording has several overlapping faults.

What sensitivity should I use for Mouth De-click?

There is no universal number. Start low on one difficult phrase and increase only until the target click is detected. Monitor the removed signal if the tool offers that output; back off when it contains consonants, breaths or vocal texture. Always compare normal repaired speech with the source.

Can AI remove mouth clicks automatically?

Some AI voice cleaners advertise mouth-noise removal, but the same safety rule applies: compare with an untouched source at matched level and inspect word endings. A generic speech enhancer is not automatically a specialist de-clicker. Before uploading unreleased client dialogue to a web tool, check its storage, retention and model-training terms.