COMPLETE GUIDE
iZotope RX Dialogue Cleanup: A Module-by-Module Breakdown
Module-by-module walkthrough of dialogue cleanup in iZotope RX: Voice De-noise, De-hum, De-reverb, Mouth De-click, Spectral Repair and De-wind. What each module does, which controls it has, and where it stops working — based on the iZotope RX 11 documentation.
KKirillFounder & Creative Director6
RX modules covered
RX 11
Version the manual describes
50 Hz
Mains hum frequency
iZotope RX is an application for dialogue cleanup: spectral editing, noise reduction and targeted repair modules in one place. This guide covers the modules that matter for speech clarity: Voice De-noise, De-hum, De-reverb, Mouth De-click, Spectral Repair and De-wind. The description follows the iZotope RX 11 manual: each module has its own controls and its own limits, and below is what the manufacturer documents.
Voice De-noise
Adaptive vs Manual Noise Reduction
Voice De-noise removes broadband noise from speech recordings: hiss, aircon rumble, fan noise, room tone. Under the hood there are 64 psychoacoustically spaced bandpass filters acting as a multiband gate: what sits above the threshold passes unprocessed, what sits below it is attenuated.
Learn mode (manual). The noise threshold is set from a pure-noise reference: select a passage that contains only background and click Learn. The manual recommends at least 1 second of pure noise — longer selections set the threshold nodes more accurately.
Adaptive mode. The module adjusts the threshold automatically as the noise floor changes; in Adaptive mode the threshold is deliberately set lower so that artifacts do not appear. This mode suits recordings with a changing background.
The threshold is a curve of 6 nodes: set them by hand, or let them follow automatically in real time when Adaptive Mode is on.
Filter Type offers two modes: Surgical removes more unwanted noise but sacrifices timbre and can produce musical-noise artifacts; Gentle is more transparent but removes less high-end sizzle. Dialogue and music have separate modes: Optimize for Dialogue reacts faster to noise changes but is not meant for noise in sung vocals, while Optimize for Music does not attenuate sustained notes.
Voice De-noise Best Practices
- Always learn the noise profile from a passage of clean background: in manual mode the threshold comes from it, and the whole chain depends on that passage being clean.
- In Adaptive mode the threshold follows by itself, but it is deliberately set lower — fewer artifacts, and therefore less noise you can cut.
- Listen to the difference: the Difference output shows what RX removes. If you can hear speech in it, the threshold is too low.
- Move the threshold nodes by hand: there are 6 of them, and in Adaptive mode they set themselves.
- Dialogue and music have different modes: Optimize for Dialogue reacts faster to noise changes, Optimize for Music does not attenuate sustained notes.
De-hum: Mains Hum
Removing Mains Hum
Mains hum is a constant background from the power grid: grounding problems, cheap power supplies, cables running close to the microphone. Its frequency is 60Hz in America and 50Hz in other parts of the world.
The De-hum module has 2 modes. Static is for simple tonal hum with a few harmonics, where the hum does not overlap speech frequencies. Dynamic is for complex hum: more than 12 harmonics, inharmonic hum, buzz reaching into the upper frequencies, or hum that overlaps speech.
Controls in Static mode: Base Frequency sets the fundamental of the hum, found by ear with Preview while the spectrum analyzer helps you; Filter Q sets the bandwidth of the notch filters for the fundamental and its harmonics; Learn sets the base frequency automatically from a selection containing the hum; Adaptive Mode adjusts the notch filters over time when the hum changes pitch.
Dynamic mode uses a different set: Sensitivity controls how much hum is removed; Bands is the number of dynamic notch filters — raise it for complex hum at the cost of transparency; Filter Q and Range Selectors set the filter bandwidth and the bounds of the processed region.
Honest limitation: De-hum removes the hum frequency but cannot reconstruct dialogue that was distorted by overload. If the hum is loud enough to flatten the waveform, the signal is damaged and the module cannot bring it back.
De-hum Troubleshooting: When Hum Persists
If the hum does not fully go away, work through this: — Is the base frequency set correctly? Some equipment hums at frequencies slightly offset from the mains value, so set Base Frequency from the peak on the spectrum analyzer (Window > Spectrum Analyzer). — Are enough harmonics covered? For complex hum switch to Dynamic mode: it is designed for hum with many harmonics, for inharmonic hum, and for sound reaching into the upper frequencies. — Is there broadband noise under the hum? If there is hiss or fan noise as well, De-hum alone is not enough: lower the broadband noise with Voice De-noise first, then return to the hum. — Is the hum getting in the way of noise reduction? Blackmagic Design states this directly: when a clip contains both hum and noise, remove the hum first so the noise reduction works on noise that sits in a different frequency range.
De-reverb
Reducing Room Echo and Reverb
De-reverb reduces the room captured by the microphone: reflections and the reverb tail. The module works on the reverberant-to-direct ratio (wet/dry). Learn analyses the signal and determines that ratio per frequency together with the overall decay rate, then sets Reverb Profile and Tail Length itself. The manual recommends learning on a selection that starts with some noise floor, runs for several seconds, and contains both the direct sound and its reverberant tails.
Tail Length approximates RT-60: the time it takes for a reverberant signal to fall by 60 dB in amplitude. If tails reappear after processing, raise it; if the result sounds dull and over-processed, lower it. Early reflections — the fast echoes of a direct sound from a nearby surface — typically occupy the first 5–100 ms of the tail, and the minimum Tail Length setting is what works against early reflections.
Artifact Smoothing controls frequency accuracy: it is very high by default because reverb is smooth across the spectrum, but lowering it handles resonant tones more accurately and adds artifacts. Enhance Dry Signal raises the level of the direct signal, which restores dynamic range and helps prepare the material for later noise reduction. Output Reverb Only leaves only the wet reverberant signal at the output, which makes it easy to hear what the module considers reverb.
How to apply it: De-reverb ships as a VST/AU/RTAS/AAX real-time plug-in, but the processing is complex and resource-heavy. For high-quality results, open the file in the RX Audio Editor, apply the module and return the result to your session.
Mouth De-click
Removing Lip Smacks and Mouth Clicks
Mouth De-click removes mouth noises: tongue clicks, lip smacks, wet mouth sounds. The module is designed for longer selections, but it can also take out individual clicks.
Sensitivity determines how many clicks are detected; raise it too far and the module starts hitting plosive consonants and damaging the original signal. Frequency Skew targets click detection at lower or higher frequencies: negative values suit generic clicks such as those on vinyl, while zero and positive values target mouth clicks in the middle frequencies. Click Widening extends the repair area around a detected click, which matters for sounds with a decay such as lip smacks.
Workflow: run the module across the whole dialogue track, listen, and repeat where needed. Running Mouth De-click twice sometimes produces a better result than a single pass, because the first pass can miss quieter clicks masked by louder ones.
Tip: run Voice De-noise first, Mouth De-click second. On a high noise floor the clicks disappear into the noise and detection becomes less accurate.
Spectral Repair
Manual Frequency Repair for Isolated Noises
Spectral Repair is manual repair in the spectrogram: you select the offending region and the module rebuilds the sound from the surrounding data. It treats the selection as corrupted audio and repairs it from the material around it.
Three modes. Attenuate compares the content inside the selection with the content outside it and reduces the level of the selected area to match the surroundings without resynthesising audio — this is the mode for background noise and for events that do not fully mask the signal you want. Replace substitutes the selection entirely with audio interpolated from the surrounding data, for badly damaged passages of tonal sound. Pattern matches a repeating pattern from neighbouring regions.
Controls: Bands sets the number of frequency bands used for interpolation — more bands give better frequency resolution but need a wider surrounding area, and short selections or transients need fewer; Surrounding Region Length sets how much surrounding sound is taken for interpolation; Multi-Resolution separates frequency resolution for low content from time resolution for high content; Direction of Interpolation chooses where the material comes from — signal to the left and right, above and below, or in all directions.
Honest limitation: when noise overlaps speech in frequency and time, the repair takes the speech with it. Spectral Repair can reduce the noise but cannot choose between the noise and the words — that stays with the operator.
De-wind
Removing Wind Noise from Outdoor Dialogue
De-wind removes the intermittent low-frequency rumble that gusts of wind produce when they hit the microphone diaphragm. The module is built for light and moderate gusts; it does not handle heavy wind that blows out and distorts the microphone signal.
Controls: Reduction sets the balance between the depth of wind reduction and the preservation of the original signal; Crossover Frequency is the upper frequency limit of the algorithm; Fundamental Recovery re-synthesises lower voice harmonics that may be lost under the wind; Artifact Smoothing removes the musical noise characteristic of FFT-based processing, which sounds underwater.
Important limitation by scenario: the module suits intermittent foreground wind — a recording with periodic gusts hitting the diaphragm directly. If the wind is constant or in the background, De-wind treats it as part of the noise floor and preserves it. For such recordings the manual suggests Spectral De-noise or another pass of De-wind.
Typical Scenarios
Hotel Interview with Aircon and Mains Hum
Scenario: an interview recorded in a hotel conference room. The aircon gives a constant low-frequency rumble, a grounding problem adds mains hum, the tile floor adds reflections, and the lavalier microphone picks up mouth clicks.
Processing chain:
— Voice De-noise in Learn mode: learn from the pause between lines, where only room tone is audible.
— De-hum: set the base frequency of the hum; switch to Dynamic mode if the hum is complex.
— De-reverb: learn on a passage that contains the speech and its tail, and raise the direct signal if needed.
— Mouth De-click: run across the whole track and repeat where needed.
Result: the dialogue is clean, present and natural; residual room tone is barely audible.
Street Vlog with Motorbike Traffic
Scenario: a presenter speaking straight to camera on a busy city street. Passing motorbikes overlap the speech in frequency and time, so they cannot be cut out individually; constant city noise stays as the bed.
Processing chain:
— Voice De-noise in Adaptive mode: the noise changes constantly and the threshold follows on its own.
— Spectral Repair on the isolated horn honks between words: the region is selected in the spectrogram and replaced with interpolation.
— De-hum: the base frequency of hum from the nearby transformer.
Result: the presenter's voice is clearer and the constant bed is reduced, but the motorbike passes are still audible — they cannot go without damaging the speech. For a clean result, re-record the voice in a studio and sync it to picture.
Beach Wedding with Coastal Wind
Scenario: a wedding ceremony on a windy beach. Steady coastal wind rumbles in the officiant's lapel mic, gusts cause momentary overloads, and guest chatter shows up in the quiet passages.
Processing chain:
— De-wind: removes the steady wind rumble; a second pass helps with a constant bed.
— Voice De-noise in Learn mode: learn from a quiet passage where only guests are audible.
— Spectral Repair on the clipped gust moments: the selection is replaced with interpolation and some speech is inevitably lost there.
— De-reverb: outdoors there is little reverb, while the open air adds a slight diffuse quality.
Result: the ceremony audio is usable; the clipped moments lose some quality but the words stay intelligible.
Processing Order and Workflow
Recommended Module Processing Order
The order in which you apply modules matters: earlier modules can interfere with later ones.
There is one rule the documentation confirms: hum first, then noise. If a recording has both mains hum and broadband noise, remove the hum first, so the noise reduction works on noise that sits in a different frequency range instead of fighting the hum.
Beyond that, pick the order for the recording. Isolated events (door slams, phone rings, equipment clicks) are convenient to remove with spectral repair before noise reduction, so the automatic modules do not latch onto them. Remove wind rumble before noise reduction as well, because it also raises the noise floor. Reverb and mouth clicks go last, when the signal is already clean.
This order is a starting point, not a rulebook. Do not apply a module to a problem the recording does not have: every processing step costs something.
RX Workflow Best Practices
- Work on short sections at a time, not the whole file: it is easier to A/B compare before and after and to set parameters per section.
- Always keep the original file: RX modifies audio in place, so work through File > Save As.
- The History panel (Window > History) undoes any processing step and returns you to any point in the chain.
- Render and listen after every module: several modules in a row will not tell you which one caused the artifact.
- Compare against the original often: after several modules your ear adapts and you lose perspective.
- Export the difference (Module > Output: Difference) and listen to it: if you hear speech, you removed too much.
iZotope RX Dialogue Cleanup FAQ
How much noise reduction is too much in iZotope RX?
The documentation gives no numeric limit. It says something else: in Adaptive mode the threshold is deliberately set lower so artifacts do not appear, and Surgical mode removes more at the cost of timbre. If you hear artifacts, raise the threshold or use the gentler filter type; if noise remains, lower it. Noise this heavy is better fixed by re-recording than by squeezing the reduction.
How do I remove 50Hz hum from location audio?
With the De-hum module. Mains hum runs at 50Hz across Europe and most of Asia, and at 60Hz in America. Set Base Frequency from the peak on the spectrum analyzer (Window > Spectrum Analyzer) and keep Filter Q narrow, so you cut the frequency without touching the surrounding sound. If the hum is complex, switch to Dynamic mode. If the hum stays, check whether there is broadband noise next to it: then run Voice De-noise first and De-hum after.
Can De-reverb completely remove room echo from dialogue?
Not completely: the module works on the reverberant-to-direct ratio and only reduces the share of reverb. The documentation does not put a number on how deep it can cut. If tails reappear after processing, raise Tail Length; if the signal sounds dull and over-processed, lower it. For severe reverb, re-recording the speech in a treated room is the practical answer.
What is the correct processing order for RX modules?
The documentation confirms one order: hum first, then noise. Blackmagic Design states it directly: when a clip contains both hum and noise, remove the hum first so the noise reduction can focus on the noise, which is likely in a different frequency range. The rest is your call: put De-reverb and Mouth De-click where they earn their place, and never apply a module to a problem the recording does not have.
Can iZotope RX fix clipped or distorted audio?
Partly. The De-clip module repairs clipping artifacts by interpolating the waveform above a given threshold, so it helps where clipping has slightly flattened the peaks. Set the threshold just below the level where the signal clips. If the signal was overloaded badly enough that the distortion is in the file, the repair cannot bring back the original waveform — re-recording is more reliable.
How do I use RX with Pro Tools?
Through the RX Connect plug-in: it sends a clip or several clips to the standalone RX 11 application, where you work in the spectrogram, and in Send For Repair mode it returns the processed result to the host. In Pro Tools the plug-in is available from the AudioSuite menu. The alternative is to export clips, process them in standalone RX and bring them back into the project.
Need Professional Dialogue Cleanup?
Our audio engineers clean dialogue recorded in challenging environments — aircon rumble, busy street traffic, gusty sea wind, tile-room echo. We use iZotope RX, Pro Tools, and location recording experience. Send us a sample and we will assess the audio.
Get a Free Audio Assessment