iZotope RX

There is a category of audio problem that no amount of skill fixes with conventional tools, and everyone who has worked in post knows the feeling.

A performance you cannot get again, with a fault printed into it: a lavalier rustling against a costume, a fridge compressor, a plane, a chair creak on the one take where the actor found it.

iZotope RX exists for exactly that, and it is the reason the software is close to standard equipment in post-production. RX 12 is the current version, and the interesting question is not whether it works but which of its fifty-odd modules you actually reach for.

RX 12 In Five Points

• More than 50 tools across restoration, dialogue rescue and sound design, aimed at film, TV, music, podcasts and content production.
• Neural-network modules do the hard separations: Dialogue Isolate, Music Rebalance, Breath Control and Repair Assistant all run on new neural nets in RX 12.
• Scene Rebalance and Trim Silence are new this version, alongside rebuilt De-bleed and Breath Control.
• Four editions — Elements, Standard, Advanced, and RX Post Production Suite 9 — with the module count rising at each step.
• It works as a standalone editor and as plugins, including Pro Tools AudioSuite, so repairs can happen in the session or outside it.

Why Does Spectral Editing Beat A Better EQ?

The idea underneath RX is that audio can be edited as a picture rather than as a signal, and understanding that explains why it recovers material other tools cannot.

A conventional equaliser works in one dimension. You attenuate a frequency band, and you attenuate it for the entire duration of the audio — which is fine for a constant hum and useless for a cough, because the cough and the voice occupy the same frequencies at different moments.

A spectrogram adds the second dimension. It plots frequency against time, so a cough is a shape you can select and remove while leaving the words either side of it untouched. Spectral Repair then interpolates across the hole from the surrounding material, which is closer to content-aware fill in a photo editor than to any audio process.

That is the conceptual jump, and it is why RX's interface looks unlike a mixing console. You are not adjusting a signal path; you are selecting regions of a picture and telling the software what to do with them.

The practical consequence for the fifty-plus modules is that most of them are automated versions of a selection you could make by hand. De-click finds and repairs the clicks; De-rustle finds the clothing noise; Mouth De-click finds the specific sound of a dry mouth opening. The value is in knowing which one matches the fault in front of you.

This is also the line between RX and the one-click enhancers. A tool such as Adobe Podcast makes one decision on your behalf and applies it to everything; RX hands you the decisions and assumes you know which fault you are looking at.

A spectrogram turns a cough into a shape you can select and remove, leaving the words either side untouched. That is not filtering — it is editing a picture.

What The Neural Modules Do Differently

RX 12 states brand-new neural nets behind Music Rebalance, Dialogue Isolate, Breath Control and Repair Assistant, and these are a different class of tool from the spectral editors.

Source separation is the capability. Dialogue Isolate takes a mixed recording and pulls the speech away from everything else, and Music Rebalance takes a finished stereo mix and separates it into vocals, bass, percussion and other — from a mix, with no access to the multitrack.

The site describes this as reference-free source separation, which is the honest framing. The model has learned what voices and instruments sound like in general, and it uses that knowledge to guess which parts of a mixed signal belong to which source.

That guess is remarkable and it is a guess. Separation quality varies enormously with the material: a clean voice over steady traffic separates beautifully, while a voice competing with another voice at the same level does not. Nothing in the interface tells you which case you have until you listen.

The practical discipline is to audition the separated stems in isolation before committing. Artefacts that are inaudible under a full mix become obvious the moment the mix is removed, and it is far cheaper to discover that before you have built a session on the result.

Choosing Between Elements, Standard And Advanced

Four editions exist — RX 12 Elements, Standard, Advanced, and RX Post Production Suite 9 — and the difference between them is module count rather than quality.

The way to choose is to work backwards from the faults you actually encounter, because the expensive modules are the specialised ones. If your work is podcasts, the tools you will use constantly are De-noise, De-click, Mouth De-click, De-plosive and De-ess, and those sit low in the range.

Location dialogue is what pushes you upward. De-rustle for clothing noise, De-wind, De-bleed for microphone spill between characters, Dialogue Isolate for the aeroplane, Ambience Match so a repaired region does not sit in a different room from the material either side — that is the set that justifies Advanced.

Ambience Match is the one people underestimate. Remove noise from a single word and you have created a hole in the room tone, which the ear notices immediately as a discontinuity. Matching the ambience back is what makes a repair invisible rather than merely clean.

Music Rebalance is the outlier. If you ever need to lift a vocal out of a finished mix — for a documentary, a remix, or a cue where only the stereo master survives — nothing else in this suite does it, and no cheaper tool does it as well.

For podcast work specifically, an automated cleaner such as Cleanvoice AI handles filler words and mouth noise at volume, and the honest comparison is that it is faster at the routine cases while RX is the one that rescues the episode with a builder outside the window.

• Match the edition to your faults, not to the module count.
• Podcast and interview work rarely needs the location-dialogue modules.
• Location dialogue is what justifies Advanced: De-rustle, De-wind, De-bleed, Dialogue Isolate, Ambience Match.
• Ambience Match is what makes a repair invisible rather than obviously repaired.
• Music Rebalance is a capability, not a convenience — nothing cheaper replaces it.

Every Repair Is A Trade, And The Trade Is Artefacts

This is the part that separates people who own RX from people who are good with it, and it is worth stating before any module is opened.

Noise reduction does not remove noise. It attenuates whatever the algorithm believes is noise, and everything it removes takes some of the wanted signal with it. Push it far enough and dialogue acquires that watery, underwater quality that instantly identifies over-processed audio to anyone who has heard it before.

The correct target is almost never a clean recording. It is a recording where the fault is no longer distracting, which is a much lower bar and a far better-sounding result.

That means working in small amounts across several modules rather than a large amount in one. Six decibels of broadband reduction, a gentle de-hum and a de-click will beat eighteen decibels of de-noise every time, because each process is operating in the range where its artefacts stay below audibility.

It also means A/B testing against the original constantly, at realistic volume. Processing decisions made at high monitoring levels on headphones routinely sound wrong on a laptop speaker, which is where most of the audience will hear it.

The same trade governs real-time suppression. A live tool like Krisp has to decide instantly and cannot be auditioned, which is why its results are acceptable on a call and not acceptable on a delivery master.

A Working Order For Repairing A Damaged Take

Order matters more in restoration than in almost any other audio work, because each process changes what the next one sees.

De-noising first, for instance, leaves clicks sitting in near-silence where they become more audible rather than less.

1. Listen to the whole take first and write down every distinct fault, since the tool you need depends on naming the problem correctly.
2. Fix the impulsive events first — clicks, plosives, mouth noise — because they are localised and removing them does not affect the rest.
3. Remove tonal problems next, such as hum and whine, which are narrow and cheap to attenuate.
4. Only then address broadband noise, and take a noise print from a genuinely empty section rather than a quiet one.
5. Apply Ambience Match after any repair that leaves a hole in the room tone, so the fix does not announce itself.
6. Compare against the unprocessed original at a realistic listening level before rendering, and if the processed version sounds impressive rather than natural, do less.

iZotope RX 12: Editions And Modules