Audio Repair 101: Ambient Noise

Get access to a comprehensive guide on noise reduction by Emmy®-winning dialogue editor John Purcell. Learn how to control ambient sounds like wind, traffic, and electrical hum that run the length of a scene, whether they’re broadband or tonal.

This is written from the perspective of film dialogue postproduction, but noise is noise, so what you learn can be applied to other audio recordings as well.

What is ambient noise?

Ambient noises are the unwanted sounds that run the length of a scene. They may be steady, they may change over time, but they are generated by the space and its surroundings rather than by short events within the space. This article will help you control these noises.

There are two basic categories of ambient noise: harmonic (tonal) and broadband. Today you can probably use one plugin to process a track that contains both harmonic and broadband problems and get decent results, but usually you should try to do your ambient processing in two steps. This lets the processors do what they do best, and keeps you from having to overcook the tracks.

Reducing harmonic noise

Harmonic noise is, by definition, harmonic, which means that in most cases you clean it up with a multiband EQ.

Use a multiband EQ to create a custom harmonic filter. Locate and attenuate the fundamental sound and as many of its harmonics that it takes to control the noise. Notice that in the image below, the Q and gain cut are tailored for each band. This enables you to find the greatest harmonic reduction per band with the least damage to the dialogue.

The process is easy:

• Pick an EQ with many bands: one that will let you make extreme settings in gain and Q. Generally, EQs that are very “musical” are inappropriate, since they won’t let you make super-deep cuts with very, very high Q settings. The Waves Q10 is a classic for this kind of work: it lets you do the crazy things that are needed to control harmonic ambient noises.
• Use the second band to locate the fundamental (save the first band for a high pass, should you later need it). Raise the gain all the way and set the Q pretty high. Starting at a very low frequency, move the frequency back and forth until you hear a loud low-frequency rumble (left). This is the fundamental—you’ll know it when you find it.

• If you hear that you’re cutting into the dialogue in any way, reduce the cut, change the Q, or check if the frequency is wrong. As you see in the images above, the gain does not need to be a complete notch, and Q does not have to be razor thin. Like everything in noise reduction, it’s all a compromise.
• Repeat the process for at least five harmonics, preferably more. Listen with each band on, then off. Also, bypass the whole plugin occasionally, always remembering that it’s the voice that matters, not so much the lack of noise. Be diligent and honest each time you compare the processed sound with the original. If you can’t make this work, then try another route. Make presets as you go, since you’ll undoubtedly need to get back to these settings.
• Save at least one band for frequencies around 8 kHz. You’ll almost always find some “cricket/tinnitus” mess up there. A gentle notch here really saves your nerves.
• Don’t use a high-pass filter as a primary tool against harmonic noise. If you raise its frequency enough to clean a decent amount of noise, you’re likely gelding the voice. However, as you see in the example above, a conservative high-pass will likely remove some of the inarticulate muck at the very bottom.

Remember, listen. Solo and/or mute each band each time you change a setting. Bypass the plugin and compare with the original. Listen past the noise. Stick your head through it and hear the voices of the original. Compare. Listen, listen, listen.

Traps to listen for when using an EQ for tonal noise:

• Voice damage in certain frequency bands. Be straight with yourself and back off or turn off the band.
• “Singing” EQ bands. Q is too high.
• No apparent difference between the processed and original sounds. If a band doesn’t appear to be doing anything, turn it off. If it sounds like you aren’t getting anywhere overall, stop.

The “Full Reset” menu item in the WaveSystem Toolbar resets the plugin to its factory settings. Sometimes it’s better to give up and start from scratch than to keep fixing a fix of a fix. You’ll likely come up with better results when you’re not burdened with previous misjudgments.

Pure hum

Background noise that’s purely electrical in nature can be reduced or eliminated with a hum filter. These procesors create a notch pattern of based on a harmonic series based on a selected fundamental. Typically, these devices default to 50 Hz or 60 Hz, for obvious reasons.

X-Hum plugin

The cleaner the contaminating noise, the better chance you stand of reducing it. Artifacts in the signal, reflections, and other imperfections in the noise will make the processor less effective.

Reducing broadband noise

Reducing broadband noise is more complicated and more prone to abuse than using an EQ to remove harmonic noises. It’s so powerful and so fun that you can easily be seduced into overdoing it. Take your time. Temper your enthusiasm. Know your limits. Listen.

Broadband noise reduction works by creating a description of pure noise: noise that’s not contaminated by words or other sounds that relate to the dialogue. Some denoisers require a user-created sample of pure sound. These plugins are used primarily for offline work. Other processors extract sample noise from the signal in real time. They also have little or no latency, so they’re well suited for mixing.

Broadband basics

All broadband processors work in more or less the same way. Here’s a simplified description of what they do.

1. They receive a sample of noise, transform it from the time domain (that’s where we live) to the frequency domain (a world that consists only of frequency and amplitude, and is probably a bad place to live), and then describe it mathematically. This frequency spectrum is divided into lots of very tiny frequency bins (bands), anywhere from a few hundred to a few thousand. The resulting formula is the noise profile (different manufacturers call it different things).
2. The input signal is played into the processor and transformed to the frequency domain. At each of the tiny frequency bins, the plugin compares the noise profile with the input signal. If the two descriptons are sufficiently similar, the processor declares the signal at that frequency, at that moment, to be noise. That narrow band will be attenuated by a user-determined amount, determined by “Threshold” and “Reduction,” or something similar. This process is repeated for all of the frequency bins. At frequencies where the noise profile and the input signal are not very similar, the signal passes unscathed: it is not noise as far as the processor is concerned.
3. Once this selective attenuation is complete, the bins are recombined and returned to the time domain. The slope of this recombination influences the final sound: the steeper the slope between adjacent frequency bins, the sharper the final processed sound (usually more effective but potentially more artifact prone). Smoother slopes will give you a signal that’s more comfortable, but likely a bit less quiet. Most devices have a control that let you control this choice. The options are pretty bipolar: Smooth/Punch; High/Low; Music/Gate; Warmth; Smoothing; etc.

Offline broadband processing

Each plugin has its own types and names of controls, and the details of operation differ a bit, so this example is only a basic tour of the essentials. For proper instructions, read the plugin’s user guide (always a good idea!). In this section we’ll look at three different broadband reduction plugins. They represent three different generations of broadband processing. If you understand the basics of these three plugins, you’ll grasp the workings of all of them, regardless the manufacturer.

X-Noise is a classic broadband processor, where you manually locate a chunk of room tone that’s as free from desirable signal as you can find. The plugin then makes a noise profile. You know the rest.

Here’s what you need to know to run this kind of processor:

1. Select bit of clean room tone (see example above). Click Learn. In this mode, input signal (your selected pure noise) is used to calculate the noise sample. Play audio (since we’re processing offline, press Preview on the plugin). You’ll see the white Profile line moving about. Once you’ve played a good snippet of sound, click Learn once again, and the white line freezes. You’ve got your noise sample.
2. Select the section of audio you wish to process.
3. Preview the input signal. Adjust Threshold and Reduction. These two controls are very interactive, so experiment with both. You can quickly start creating artifacts if your reduction is too high. Back off.
4. Use the Resolution buttons to set how the signal moves from the frequency domain to the time domain (see #3 in the previous example). This is where you set the recombination characteristics.

The other X-Noise controls have to do with attack and release for the gate-like attributes, and a high shelf. Read the manual.

This generation of broadband processors provides miraculous tools for reducing noise without undue damage to high-frequency sounds. They are still widely used. But soon an additional need arose.

X-Noise and its brethren reduce broadband backgrounds using a noise profile and two basic controls. This is effective much of the time, but when Threshold and Reduction are the only controls at your disposal, you at times have to choose between adequate cleaning with some artifacts or no artifacts but less noise reduction than you’d like. The solution is to manipulate the curve of the noise sample itself.

Z-Noise adds a noise profile editing tool (the diamond-shaped makers on the graph). It lets you, in essence, re-write parts of the noise profile. You can change what the processor chooses to attack and what to let pass.

By reducing the amount of model-based attenuation at a specific frequency (move the marker down), those frequencies are processed less (probably resulting in fewer artifacts and perhaps more noise at that frequency only).
When you lower a marker, the associated freqenency will likely get louder.

Something else that developed in the generation between X-Noise and Z-Noise is the ability to extract noise profile from a signal directly. Sometimes you just can’t get a decent bit of noise, or it keeps changing. In such cases, Extract creates an ever-changing noise sample.

Online broadband processing

Since online processors are meant for mixing conditions, a few things are expected of them:

• The ability to reduce noise when it’s not possible to take a static noise sample. A new noise profile must be made all the time.
• No latency. It’s pretty obvious why you don’t want significant latency during a mix.
• Capacity to tackle harmonic and broadband noises simultaneously. I’m a true believer in the two-step path to background noise reduction: kill off the harmonics and then do the broadband stuff. But modern postproduction is so action packed that you really need to have at your disposal processors that adequately kill off most everything that’s thrown at it.

Most online noise-reduction processors use the Z-Noise “extract the noise sample” model. You’ll typically see a bank of faders, each of which controls a processing band. Next, some sort of threshold control, and finally a Resolution control. Its range could be “Low to High,” or “Smoothing 0–100,” or “Smooth to Punch,” or any other such thing. It’s the sharpness of recombination back to the time domain.

WNS Noise Suppressor plugin

If you’re comfortable with the Z-Noise controls, the WNS Noise Suppresor will come naturally:

1. From the WaveSystem Toolbar, select a band range. This provides greater band resolution, since the row of faders represents a narrower frequency range.
2. Set all faders to zero.
3. Play sound. Aggressively move one fader at a time to learn how each band lessens background noise, and how it impacts everything else. Once you’ve listened to the impact of one fader, return it to zero (Alt+click returns any parameter control to its default positon). Always check one variable at a time. Right now you’re just getting to know how each control influences the noise and the voice. Nothing more.
4. Apply the least necessary reduction to each band. If a processor band is not doing anything helpful, reset its fader to zero. Don’t process anything you don’t need to process.
5. When you’re happy with the band settings, use the Threshold control to find the desired level at which processing begins. Set too high and processing is always engaged, possibly creating artefacts. Set it too low and you may hear pumping or there may be not noise attenuation at all. Again, it’s a compromise.
6. Save your presets.
7. When you’re setting the threshold, you may want to play the dialogue against the backgrounds and other mix elements. If you do all of your dialogue work in a vacuum, you may overprocess and cause unnecessary damage. But it you hide behind your backgrounds, you may under-process. Yet another compromise.
8. Finally, adjust the “Smooth” fader to find your ideal compromise between sharp or smooth sound.

Of course, if you’re working with inserts, you can always change the settings later. Buy why not try to get it right now, since the sound of the dialogue will affect other elements?

There’s really nothing more to know. Well, that’s not true. There are the matters of skill, experience, curiosity, and diligence that make for a good editor or mixer.

Remember: A little bit of broadband noise reduction goes a long way, so take it easy.

Working with inserts vs offline processing

Whether to work online or offline is not always obvious. Clearly, if you’re mixing, you’re likely going to process on inserts, and if you’re trying to save a scene in which each take of each shot is unique, you’re best served with offline processing. But there are some gray areas to explore.

Inserts

The upside of using inserts for noise reduction is that you work in real time: you can hear all of the processed dialogue in context, and you can easily tell if you’re doing too much/little work at a specific moment. In other words, it’s live. Inserts provide you quick, flexible solutions. Inserts are more forgiving than offline processing such as AudioSuite.

The biggest downside of insert processing in non-mixing settings is about you, not so much about processors: insert noise reduction processing can make you lazy. By the time you’re ready to work on noise reduction, you’ve already arranged the film by scenes, organized the scene’s clips by shot and other variables, found separate room tone for each shot, created sensible shot transitions that will minimize processing, removed all the transient noises that you have patience for, and basically worked yourself silly to make tracks that really work well and are very mixable. (If you don’t understand what I’m talking about, read my book on dialogue editing. I know this sounds like a promo—it is, of course—but it’s the quickest way for you to learn how this works).

Despite your well-planned tracks, you may break down and shove all the dialogue channels into one buss, slap on some generic noise reduction, and call it a day. This is a common, self-destructive mistake. Even if you do your EQ processing at the track level, and then the broadband processing on a dialogue buss, you undermine your well-orgainzed and caring dialogue edit. All the dialogue is going down one noise reduction highway, and you’ll inevitably have to over-process in order to iron out the mismatching room tones. You’ll probably say, “I’d never do something so obviously silly.” You will.

However, real-time insert processing needn’t result in a compromised, overcooked, boring scene. If you keep the bulk of your processing on the channels (not on the dialogue buss) and customize the settings for each channel in the scene, you can do the correct—and minimum—noise reduction processing for each shot. On the buss you can do whatever processing is common to all the tracks in the scene. This lets you use the dialogue buss do the make the scene relatively quiet and smooth, without having to over-process. Remember to save all of your plugin presets. If you rely on the DAW to save your plugin settings, you’ll lose a lot of flexibility down the line.

One more thing about inserts: latency. Many modern processors work with zero latency, so you can pile lots of them onto a channel with no concern of slowing down the whole channel, buss, or even mix. However, some perfectly good processors have notable latency. This makes them problematic as inserts. Yes, you can offset the channel itself to match the latency, but there’s a limit to how much of this you want to do.

Bottom line on inserts: This is a perfectly sensible way to work, by far the most common way—in fact, the only way—to mix dialogue tracks. Save your high-latency plugins for offline processing.

Offline processing

Certain situations aren’t well suited for insert processing:

• Shots with great variability of noise over the course of the scene, among different takes of a shot, or even within a take. Traffic, wind, and interior car noise are classic examples.
• Very small sections of processing, such as mic boom noise (discussed earlier).
• Situations where you want to keep track of each iteration of the process. This is especially useful when doing complex series of processing steps and you want a sensible way to get back to any previous layer.
• Times when you need to know if there’s any possible chance that you can save the scene by throwing everything you have at it. If it looks like this won’t work, you’ll have to call for ADR. This is a Hail Mary noise reduction attempt, so to speak.
• You want to streamline the mix by providing some noise-reduced options. If you get it right, you can really help the mix.

Offline noise reduction is not unusual. In some ways it’s more focused and precise, since you are hand-making each clip and you’re probably spending more time doing it. Some processors that offer insert processing provide a “higher quality” offline algorithm that delivers a quality that can’t be delivered in real time. Offline processing clearly doesn’t care about delay, so there’s no worry about latency.

If you want to do some offline processing on a scene or part of a scene, save yourself horrible headaches by paying attention to these steps.

1. Unlike inserts, which are perpetually flexible in the mix, offline noise reduction is permanent; it’s not mixing. You are dealing with some very specific issues that will take some heat off the mix. You’re not here to “make it sound good.” You’re here to tame as much noise as you reasonably can with little impact on the feeling and sound. Save the rest for the mix.
2. Before you start your noise reduction on separate clips, fully edit the sections you intend to process. Create a beginning and end for each segment—one that makes for a good transition with the dialogue before and after. Clean up whatever bad edits exist in the material you will be processing. Remove unwanted clicks, pops, and other uglies.
3. Make a copy of each of your wonderfully edited sections. Move them to other tracks (where you move them to is between you and the mixer) and mute them. The point of this is that no matter what you choose to do offline, no matter how timid or daring or weird you are with noise reduction, there’s always a way to play the mix-ready, unprocessed, totally-edited tracks. When you get to the mix and you discover that the noise reduction you did in your studio doesn’t hold up in the mix room, you’ll need a painless way to return instantly to the unprocessed but otherwise identical edit. This one will save you. This also lets you compare your current processing with the original track, something you really need to do if you want keep track of the truth. This, too, will save you.

Save your steps and presets

It’s very likely that you’ll be doing more than one step of processesing when you work offline. After all, if your noise reduction worklow were so simple you’d be working online. Since each step is built on the previous, save each iteration. This lets you (1) back up and hear where you went wrong, and (2) blaze a new path from there. If you simply blast over each step you can’t back out of the mess you built, and you’ll have to start from scratch if you run into a brick wall.

This image shows a multi-iteration de-noise process. The original clip is on top. Each step in the process is built on the one above it. Note that as number of processes increases, the “pedigree” names become longer.

Also, at each stage, save a plugin preset of what you’re doing and name it sensibly. You will need to reproduce settings. Really. It’s totally stupid to have to recreate the parameter settings each time you need the same or similar settings.

It’s easy to forget to keep and label each layer of processing, but sooner or later you will wish that you had.

Remember Why You’re Doing This

We’ve pretty well established that every noise reduction job is different. If you can’t think of a better way or working on your project, use this sequence. At every step, compare your work with the previous one and with the original.

• Listen. Decide if there’s a problem and if so, what it is.
• Reduce most of the rumble and other harmonic noises with an multiband EQ. Use deep, sharp cuts and maybe a tiny bit of a high-pass filter.
• Deal with transients.
• Listen.
• Sort out broadband noise.
• Listen.
• If there is some manageable damage, use a gentle EQ like a Renaissance EQ to pick up some of the lost lovliness of the original. Don’t depend on this too much, since our job is to lessen the noise, not to make it “nice.”
• Leave the room for a few minutes. If, after this break, you still think you’re OK, stop. Congratulations.
• Move on.

« Read the previous article in this series: Controlling transient noises: Clicks, crackles, pops

Emmy® winning sound editor John Purcell is the author of Dialogue Editing for Motion Pictures: A Guide to the Invisible Art. He writes for Waves Audio.