Join Scott Hirsch for an in-depth discussion in this video Isolating and cleaning up noise with Spectral Repair, part of Repairing and Enhancing Audio with iZotope RX 4.
- In this movie we'll use the Spectral Repair module to deal with unwanted sound events like a brake squeal and a siren in an on-street interview. Use of Spectral Repair here is ideal, since it can isolate these sounds spectrally while keeping the narrator's voice intact. And if you use it correctly, it does it all with natural results. Let's take a listen to this clip. - All right, well, after spending some time at what could only be considered the epicenter of Times Square we walked a little bit further. We were still in the middle of the action, but when we saw the huge American flag lit up...
- So, as you can hear and see in the spectrogram, there is an obvious siren moving up and down the frequency range. It looks like this swirly kind of path that's cutting through right from around 500 to 1 kHz. There's also a whole bunch of brake squeals at the beginning of the file. You can see those as the thin little lines that appear in the higher frequencies of the spectrogram. Now, historically something like a siren poses a really bad audio issue to fix, since it changes pitch over time.
And tools like traditional fixed equalizers can't follow those kind of changes. But since we can see the audio information visually in the spectrogram, we can easily isolate it by selecting it and attenuate it with Spectral Repair. Let's start with one of these brake squeals first. The brake squeals are pretty easy to identify in the upper frequencies. So I'm going to go ahead and zoom in a little bit, And I can see I'm going to try to identify and single out one of these brake squeals. So right here is a good one. See that little narrow line there? What I'd want to do is take the Time Frequency Selector tool and I can zoom in even a little bit closer.
Let's see if I can find a real good one. I mean, you see they're all over, so we'd have to eventually fix all of these. But I'm going to find a good one to isolate. So there it is right there. Okay, so I'm going to select right around that, and I can really make the selection nice and narrow. If I wanted to I can even zoom in a little bit on the frequency spectrum like that. So I can go ahead and make a really good selection. So there we go. And let's listen to this using the Play Frequency selection. We're hearing just the brake squeak there. And again, if I take the normal play head and roll it back and play the whole thing, you can hear that right there.
Okay, so that's what we're going to isolate, and we're going to try to use the Spectral Repair module to deal with that. I just want to go ahead and set the Spectral Repair to its default value. And you'll notice there's four tabs across the top of Spectral Repair. We have Attenuate, Replace, Pattern, and Partials and Noise. Let's quickly review what all these tabs do. Attenuate is the default first tab that comes up. Essentially, it will reduce the level of a sound event, like a brake squeal as we have here, or something like a microphone bump, guitar fret noise, cough or hiccup.
All kinds of sounds that might appear in your audio recordings. Then it modifies any dissimilar audio in your selection to be more like the surrounding audio. So not only does it turn down, in this case the brake squeal, it also tries to match it to surrounding audio. I'd say about 80 to 90% of the work I do in Spectral Repair, I use the Attenuate tab. Next, we have Replace and Pattern. These two tabs work a little bit differently. They first remove audio, and then look outside the selection to seamlessly fill in the gaps.
For Replace, it fills in selected content with audio interpolated from the surrounding data. This is great if you have a dropout, or badly damaged short section of audio that you need to repair. Pattern works slightly differently. It finds the most similar portion of the surrounding audio and uses this to replace the corrupted audio. The most similar audio could be up to 10 seconds. If you change the Pattern Search range you can go up to 10 seconds away from the original sound source and it will look in there to find similar audio and replace the unwanted audio with that.
If you have audio with repetitive sounds, like say a helicopter or a fast high-hat pattern, the Pattern tab can work really well, because it will find similar sounding events and try to pattern them into place. Finally, we have Partials and Noise. It's technically the advanced version of Replace mode. It restores harmonics of the audio more accurately with control over the Harmonic Sensitivity parameter. That's this parameter here. It's best used to repair musical passages where the integrity of the note and the note's vibrato need to stay intact.
In my experience using the Spectral Repair, I probably use the Partials and Noise the least. So which is right for the brake squeal? Let's try out Attenuate first. So in terms of the parameters for Attenuate, the first thing you see is the Bands. And basically Bands refers to the quality of the process. If you go in here, higher bands offer you better resolution, but the trade-off is that they need more surrounding area to be analyzed. You don't always have that luxury. Typically I usually start with just the default setting of 512 bands.
But you may need to go in here to up the bands to get higher resolution if things aren't really working for you. You've probably noticed now that there's some extra wings added on to our original selection. This is showing us the area outside our original selection that's going to be used to modify the area once the brake squeal is reduced in volume. Notice if we go over here to the direction, we can make these wings appear horizontally, like we're seeing now, where those wings appear to the right and to the left of the original selection.
Or we can choose to look vertically, where it looks above and below the original selection. And finally, we can go in here and choose 2D, where it looks both above and below, and the left and right from the original selection. I'm going to go back to horizontal for a second. And then the sliders down below surrounding Region Length, that actually changes the length of the wing. So if we go larger, we get more of an area to try to analyze outside of the original selection. And if we go less, we can choose less.
Now the reason why you'd want to do less is because, in this case, for example, we have some other audio that's appearing on the right-hand side that might not work so well to fill in the gap, right? So we might want to go less. But the other thing you can do is the before and after weighting, and that actually weights the front or back side more heavily, depending on where you put the slider. So if I go more toward the before side, notice that it's looking only at the front half, and the wing on the back side gets closed. Or I can go in the opposite direction.
So for this one, because there's more of just regular background noise in front of the brake squeal, I might choose to weight it a little bit before, so I'm grabbing just this to fill it in once the brake squeal has been reduced. And finally, we have the Strength parameter. So typically in a clip like we're seeing here, you can't really go too much above 2, because then what happens once you go up into the higher numbers, is all of the background audio will be missing from that section. So if you keep the strength, I've learned, between about 1 and 2, it'll accurately be able to fill in the background with an appropriate amount of background sound once the brake squeal is gone.
So let's try, for example, just to start out here, a strength of 1.5 and let's go ahead and process and see what we get. Okay, that's looking pretty good. I can still see the brake squeal slightly there, but let's see if we can hear it. It's actually pretty good, right? We didn't really notice it. I'll start back here. So that is actually working pretty well. Now, I want to undo that for a second and get back to the original brake squeal, and also just show you how we can work on this section with the Replace tab.
So I'm going to recreate the selection, and I'll move over to Replace tab. Now notice when I hit the Replace tab, a couple of the parameters that were in Attenuate went away. I no longer have the ability to change the strength, but I still do have the surrounding region length and the before-and-after weighting. Remember, Replace is different than Attenuate, because it's actually going to remove all audio inside my selection and attempt to replace it, interpolating from this horizontal before and after.
So again in this case I'd probably want to weight it before, because there's some other audio after. But before the clip is looking pretty solid, and I might just lower the region length so I can see there's just some good solid background noise right there. So let's try this and see if this works. So, as you can see, at least visually, it looked almost as good as the Attenuate. So, yeah, I didn't hear the brake squeal. So in that case that worked as well. So pretty typical situation where, in this case, both Attenuate and Replace worked.
And that's common, you know. You can always use the Compare button to compare different tabs and see which one's giving you the best results. Now, I'm going to undo that and I'm going to go back to Pattern, and let's just see how that works. So Pattern still has a surrounding region length, just like Attenuate and Replace, only the other parameters are a little bit different. It's going to search in a number of seconds, up to 10 seconds, around for similar sound that was in the brake squeal area. And so it works slightly differently than Replace, which is just interpolating and filling in the blank.
This is looking to pattern in the blank. And so let's just see if this works and then see what happens. I'll just keep the search range around five seconds here. See, there's no before and after weighting in this case, right, so I'll just keep the region length around 100%. So that worked pretty well too, in this case. All three of them worked. Which is, you know, sometimes that's the case. They're all doing their different processes, but they all work. Let me show you an example of a pattern where it won't work.
And that would be like this brake squeal right here. Okay, check this out. If I process this, what it's going to do is it's going to look around and fill in, and it filled in something else because it found a pattern that was outside, that in this case didn't match, right? Because if I undo, we have some other audio that's in the little wing. So in this case it's probably best to try Attenuate and we might even want to try Attenuate vertically, because you notice above and below there is some blank audio that doesn't have any other sound events in it.
So let's see if this will work better in a vertical situation for this brake squeal. Yep, sure enough it did. Let's see if we can hear it. So that took care of that brake squeal right there. So, you know, you kind of have to do trial and error. You have to go into each brake squeal, take them one at a time, make sure you get the right tab going, and eventually you'll hit on something that's going to work pretty well for all of them. But you can see it's a fair amount of work, but in the end it'll be worth it to remove all of these brake squeals, to get the cleanest recording we can.
Now I want to take one look at the big, glaring other problem in this piece of audio, which is the big siren that's happening. So here, the key to this is going to be to make a really accurate selection over the siren. For this we can use the Brush tool and the default brush size, which looks pretty good right now, but we can always change it. If we click and hold on the Brush tool, we can change the brush size. So you can make it larger or smaller. I think around 15 or maybe 17 is a good size to start with.
So now I'm going to just go ahead and try to draw in a selection that follows this siren as it's moving up and down the spectrum. So I'm doing my best freehand. And if you mess up, it's okay, you can always stop and undo it and restart. Oop, like that, messed up. So in this case, actually, I can hold Option, or Alt for Windows users, and just sort of erase that bad area. And then hold Shift and add on to my selection.
No problem, it got a little wonky there. Again, just hold Alt and erase some of that area I didn't want. Okay, so that seems like a pretty good selection around the siren. Let's take a listen to it. Again, we're hearing just what's in the frequency selection. (background ambiance) Okay, sounds like we got the siren in there. Okay, so we did hear a little bit of the guy's voice. But, you know, Spectral Repair is really good about trying to fill in only the stuff and not remove the wanted stuff, which is in this case the voice.
So we'll see how well it does. For this one we're definitely going to want the Attenuate knob. And the other thing about this is hyperselection. Being a really long selection is that horizontal or 2D aren't really going to get us anywhere, because there's nothing on the outside edges to grab from. The only one that's going to work in a long selection like this is going to be Vertical. Okay, so then I might want to narrow the surrounding region length just a little bit. I don't want to take it down all the way to the lower frequencies. I keep it pretty tight, and I'll keep the before and after weighting just at 0.
So notice in a vertical selection, the before and after weighting moves the top or bottom of a selection up and down. And we'll keep the strength at 1.5, and let's see where we get. Okay, so visually it looks pretty good. I don't see too much of the siren, but we have to always check the audio to make sure that we didn't get any of the guy's voice to sound weird by removing that siren. Let's take a listen. Let's play here both channels. - All right, well, after spending some time at what can only be considered the epicenter of Times Square, we walked a little bit further.
We were... - That's pretty good. I barely detect any siren at all, and his voice sounds pretty solid. So we'll be using even more of the Spectral Repair in our two real-world scenarios later in this course. But in general, if you make good selections like we've done here, and apply a little trial and error, maybe some comparing, you can use Spectral Repair to make your audio squeakily clean, if you're willing to put the time and effort into it.
- Exploring the workspace and workflow
- Processing files
- Cleaning audio
- Removing noise
- Working with a DAW
- Cleaning up noise in music and field audio
- Utilizing EQ Match and Ambience Match to fill gaps and match audio
- Time-stretching and pitching audio