Impulse Responses: Normalize vs. Let 'em Alone

[tommyzai]

I have a bunch of impulse responses, mostly cabs. Would there be any advantage or disadvantage to batch normalizing them to 0 dB or another level?

[CinningBao]

I would say, unless you're doing it conform to some kind of normalisation exercise you're implementing across your whole uncompressed file world for some "I'm definitely very organised" sense of accomplishment, I wouldn't bother.

Because of they way they work (which you might already appreciate) they are frequency dependent, so some a normalised impulse response might sound 'fine' when feeding it material with mostly high frequency content (for example), but might blow-up when feeding it low frequency content. Algorithmic reverbs don't suffer from this so badly.

I started going through my collection a while ago, creating presets in Space Designer, then realising I needed to adjust the Wet output level for pretty much every IR I loaded if I wanted consistency. But I've got, like, 7 million IRs, so I figured it would be quicker to just adjust then when I loaded them. And most of them needed turning down anyway, literally none of them needed turning up.

Hope that gives you some direction

[imrae]

Peak normalisation is not level matching, it seems likely this will make the experience of flipping through them worse.

[tommyzai]

Peak normalisation is not level matching, it seems likely this will make the experience of flipping through them worse.

But wouldn't normalizing make them all uniform?

[Michael L]

Peak normalisation is not level matching, it seems likely this will make the experience of flipping through them worse.

But wouldn't normalizing make them all uniform?

Pick several very different ones, make copies, normalize them and compare. I've done it

[CinningBao]

Peak normalisation is not level matching, it seems likely this will make the experience of flipping through them worse.

But wouldn't normalizing make them all uniform?

It would make them 'digitally uniform' but that isn't how we really measure these things. And it isn't really that useful when frequency content is part of the equation.

Using RMS or LUFS measurements to get them to the 'same volume' might be more useful to you. But normalisation is a dumb digital operation and shouldn't (imho) be applied carte blanche to a bunch of separate audio files, particularly IRs.

[tommyzai]

Is there any harm in normalizing IRs?

[MTorn]

I usually normalize my impulse libraries. But I’ll admit that it’s mostly from an anal retentive sense of wanting things to be tidy.
While normalization isn’t a very “smart” way of level matching, I don’t know of a more advanced alternative. And at least it deals with one common annoyance - extreme low volume outliers.

But I don’t stop at normalization. I usually clip dead silence at the beginning (I can add pre-delay later if I want it), and trim the tails once they drop off to below audible levels.
I also often remove the initial spike of the response (not for things like cabs, where I run the IR at 100%) - often that spike is too close to a copy of the original signal, and produces echoes and phasing issues when mixed with dry.

[tommyzai]

. . . But I don’t stop at normalization . . .

Ah, a kindred spirit!

[DaveClark]

Is there any harm in normalizing IRs?

There can be; depends on the IR and your material.

IRs effectively smear each sample over several successive samples. If the IR is short such as for cabs, the smearing won't usually cause overs. But if the IR is long, and your material is already normalized to 0 dB or nearly so, there is an increased likelihood of overs and/or high levels coming out of the convolution step unless you adjust the levels to compensate. Without compensation, you could end up with clipping or increased saturation, etc. downstream, depending on what else is going on.

One strategy is to just try the IR without normalization. If it's way too low, then try normalizing and adjusting to compensate. Again, with cabs you probably won't have any trouble.

When it comes to editing IRs such as MTorn described, one needs to know what one is doing. In solving the wave equation in 3-D in a room, for example, it becomes very evident that there is some exquisite timing that occurs to make everything sound right. It's really easy to mess things up.

[tommyzai]

I'm assuming that removing silence before and after the sample would be beneficial. Am I also correct to assume that the initial sample strike is important? I normally get rid (or soften) those initial clicks in other samples, but . . .

[osiris]

This is interesting because I'm using Fruity Convolver and there's a Normalize knob on it that I assume normalizes the IR. I've never used it.

[Michael L]

This is interesting because I'm using Fruity Convolver and there's a Normalize knob on it that I assume normalizes the IR. I've never used it.

There is a free VST convolution reverb that has the button to "normalize IR loudness" (vs peak normalization) sadly on by default and it certainly moderates the effect! likely because it includes dynamic range compression to avoid clipping. All converbs have gain, output, EQ, and Start so I see no need for peak normalization.

[cthonophonic]

I'm assuming that removing silence before and after the sample would be beneficial. Am I also correct to assume that the initial sample strike is important? I normally get rid (or soften) those initial clicks in other samples, but . . .

The initial transient in a reverb IR can be important if you're trying to capture the characteristics of a speaker and microphone setup along with the reverb. The dry signal might have a noticeably (or even just subtly) different character than the reflections, which are colored by the recording setup.

Another case where the initial transient can be really useful is in establishing a distance cue for an impulse that was recorded from very far away, since higher frequencies are more quickly attenuated with distance.

So basically, if you chop off or attenuate the initial transient of the IR, it could detract from the realism somewhat, but of course that may not matter if "realism" is not the goal.

This is interesting because I'm using Fruity Convolver and there's a Normalize knob on it that I assume normalizes the IR. I've never used it.

There is a free VST convolution reverb that has the button to "normalize IR loudness" (vs peak normalization) on by default and it certainly moderates the effect, likely because it includes dynamic range compression to avoid clipping.

I think there's a fair bit of inconsistency in the way that various convolution reverbs treat normalization (whether it's on or off by default, what's being normalized, etc.), which might be an argument against doing a bunch of batch processing on your whole library. Chances are, you're going to have to readjust everything to taste anyway, so it could end up being a wasted effort.

[MTorn]

With IR captures made with microphones in real spaces, you “excite” the space with some kind of brief loud noise burst. Like playback through a speaker, a burst balloon, hand clap, etc.

It’s not uncommon, especially when hobbyists make their own IRs, that you find the noise burst still present in the sample. This produces an oddly colored version of your dry signal, in addition to the actual reverberation. In those cases I always truncate the noise burst off.

Clearly there are cases where you want to keep it, like a guitar cab in a live space.

The nice thing about IRs is that they are usually regular WAV samples - you can tinker and experiment with them as much as you want.