Reverb: Is convolution the "best"??

[EnzymeX]

In the last month we have seen a bunch of raves on reverb releases like Princeton, the UAD EMT Plate, etc.

Soliciting opinions: What is "better" sonically: A good convolution plug like Voxengo with a library of good impulses, or one of the dedicated high end plugs?

FWIW, I am leaning towards convolution, but I want to make sure that is the best way to go for a first (or only) "top shelf" reverb purchase. Thanks.

[jones-y]

Matter of opinion, really.

You'll get ten people to say yea, and ten to say nay, depending on what time of the day you ask.

I will say this though, Convolution is more flexible. You can load up any audio file and use it to convolve your audio.

I have a convolution reverb, and it does me well.

[Space Boy]

Convolution processing works in the FREQUENCY domain while a typical "traditional" reverb works in the TIME domain.

Anything that can be done in the TIME domain can also be done in the FREQUENCY domain. The two domains are related by a Fourier Transform process.

The two domains are essentially equivalent from the point of view of the end result.

However, a traditional time domain reverb is typically limited to a few very specific parameters that act on the time domain signal. When the range of desired effects is limited then this can be the least CPU hungry approach.

A convolution reverb primarily operates in the Frequency domain, and because of the Fourier Transform (FFT) treatment, the effective number of parameters that can be adjusted is significantly higher (e.g. a 2048 point FFT will yield 1024 frequency amplitude and 1024 frequency phase adjustment possibilities). The extra degrees of freedom provided by the convolution reverb can either be limited to exactly mimic a traditional reverb or used completely to mimic not only reverb but also amplifier characteristics and other acoustic anomalies. However, as you might expect, this takes a lot more CPU.

In short, convolution based processing can realize any modification of the time domain signal - reverb just happens to be one single application. A time domain based reverb can just realize a reverb, and nothing else.

[Sicklecell666]

Yep..

Paul's a thread-killer..



Whatever I don't use Reaktor's Spacemaster & Laserbrew on, I'll use a Spirit canyon or Intellegent Machinery impulse response on instead.

So, in short, I like both

[jones-y]

You touched a lot of theoretical issues here. It'd take a textbook to really understand them all, but I will say that:

Convolution is carried out in the frequency domain only because it is much less processor intensive to do so. As a matter of fact, in the frequency domain, it becomes a simple multiplication. If/when we have terahertz processors in our desktops, you'll then see time domain convolution.

The extra degrees of freedom provided by the convolution reverb can either be limited to exactly mimic a traditional reverb or used completely to mimic not only reverb but also amplifier characteristics and other acoustic anomalies. However, as you might expect, this takes a lot more CPU.

Actually, IR length is directly proportional to CPU intensity, so a shorter IR which doesn't involve echoes of the original signal, such as a guitar cab IR, is actually much less CPU intensive.

In short, convolution based processing can realize any modification of the time domain signal - reverb just happens to be one single application.

Almost. Convolution can model any linear, time invariant system. In nature, there aren't many LTI systems. Many systems, like a reverberant room, exibit a LTI range of operation, however. And in some cases, like the reverberant room, non-linearities can be safely ignored.

That's the reason why Lexicon IRs sound flat, because there are non-linear processes happening inside those boxes.

Convolution is one of the two fundamental theories/processes of DSP (the other is fourier theory). Most DSP technology is founded on these two principles. As a matter of fact, convolution is the basis of the FIR filter. So you owe your EQs, filter banks, ADCs, DACs, etc. to convolution.

[jones-y]

dbl post

[Lunch Money]

I love convolution, but I'm getting more into it for things like Spectral Relativity. However, I'm finding that although convolution is only limited in power by your impulses, it becomes a chore to design and select a proper reverb from scratch, since you have to go through your library to find the right impulse as a starting point.

With 'traditional' reverb (I use that word loosely) I've not personally used anything that matches SpinAudio's Roomverb M2. That doesn't mean it doesn't exist, but if it's something THAT far out of my price range, I don't even bother demo-ing it. Roomverb is inexpensive (though not super-cheap) and the results are stunning.

Greg

[Space Boy]

Almost. Convolution can model any linear, time invariant system. In nature, there aren't many LTI systems. Many systems, like a reverberant room, exibit a LTI range of operation, however. And in some cases, like the reverberant room, non-linearities can be safely ignored.

You could use a different convolution kernel depending on spectral intensity (signal amplitude). This could realize a non-linear behavior.

Convolution is one of the two fundamental theories/processes of DSP (the other is fourier theory). Most DSP technology is founded on these two principles. As a matter of fact, convolution is the basis of the FIR filter. So you owe your EQs, filter banks, ADCs, DACs, etc. to convolution.

Depends where you come to DSP from. In Physics, it's "all" either time domain or frequency domain, where frequency domain is just a more compact/convenient way of describing what happens in the time domain.

In audio processing, Convolution is usually a time domain process which, in practical DSP, is more efficiently and conveniently realized by switching to the frequency domain. Why waste a more powerful CPU on a time domain convolution? The achievable output would be no more "hi-fidelity" than the frequency domain process. It would be better to use all that CPU power to make more sophisticated frequency domain based convolutions - don't you think?

Isn't convolution reverb just a sophisticated FIR filter anyway? The file you load into the convolution engine is a frequency domain representation of the (Finite) Impulse Response of the room you are trying to model.

[foosnark]

I've not used convo much just because SIR isn't zero-latency yet. I put it on one throwaway track I wrote to add some grungy skanky noise in the background when I didn't care that it was out of sync.

I expect when the next SIR is released -- or I acquire the means to actually buy one of the commercial reverbs -- I'll be using it a lot more (particularly with weird impulses).

[Space Boy]

I love convolution, but I'm getting more into it for things like Spectral Relativity. However, I'm finding that although convolution is only limited in power by your impulses, it becomes a chore to design and select a proper reverb from scratch, since you have to go through your library to find the right impulse as a starting point.

The library searching can be a drag. However, I suspect that it should be possible, to design a traditional reverb type user interface, that simply generates the appropriate impulse response function according to the settings of the traditionally understood parameters (Now there's an idea). Each time a parameter is changed, a new IR is generated and written into the convolution engine. The best of both worlds.

[Sicklecell666]

The library searching can be a drag. However, I suspect that it should be possible, to design a traditional reverb type user interface, that simply generates the appropriate impulse response function according to the settings of the traditionally understood parameters (Now there's an idea). Each time a parameter is changed, a new IR is generated and written into the convolution engine. The best of both worlds.

Sounds good.

Now code the bhaaastard

[I hate browsing IR's as well, but it's not too painful using SIR as it displays all IR's in a given folder..]

[Lunch Money]

The library searching can be a drag. However, I suspect that it should be possible, to design a traditional reverb type user interface, that simply generates the appropriate impulse response function according to the settings of the traditionally understood parameters (Now there's an idea). Each time a parameter is changed, a new IR is generated and written into the convolution engine. The best of both worlds.

Sounds good.

Now code the bhaaastard

[I hate browsing IR's as well, but it's not too painful using SIR as it displays all IR's in a given folder..]

Heh, that's what *I* was gonna tell him to do. Yes, it's a great idea.

Sicklecell: Yep, SIR makes it a touch easier.

Greg

[dharmawan]

The library searching can be a drag. However, I suspect that it should be possible, to design a traditional reverb type user interface, that simply generates the appropriate impulse response function according to the settings of the traditionally understood parameters (Now there's an idea). Each time a parameter is changed, a new IR is generated and written into the convolution engine. The best of both worlds.

u mean a bit like this

http://www.voxengo.com/imodeler/

or this

http://products.prosoniq.com/cgi-bin/re ... l&refno=44

[papawillow]

I like the sound of the Princeton,,, I might have to do a lot of knob tweaking or find better IR's for Pristine Space to sound as good. I tried the Noisevault IR's,, but cannot get it to sound as 'integrated'.

[kritikon]

Personally I'm happy with convolution and a good set of 24bit impulses. I would rather use native plugins, such as the Princeton (and EMT plate if it wasn't on a UAD card) purely for the extra tweakability they give you. That level of tweaking is difficult to do with convolution.

The downside to standard plugin reverbs is that none of them do the full spectrum - you won't get top drawer halls, for example, out of the Princeton, purely because it's based on the Eventide room simulator. And obviously the EMT Plate will give you good plates...but not all other types. So you need to buy several plugins to have a full range.

The upside is less CPU and more tweakability within their range of reverbs.

Maybe a year ago, I would have had to seriously think about getting the other plugins because of CPU useage. Now CPU is not really an issue for me...I only have a 2.5GHz CPU (but with 1G RAM) and SIR with even long reverbs doesn't make me watch the CPU meter any more. I can easily use 2 or 3 instance of it and leave lots of room for other plugins.

Upside of convolution is you're only limited by the quality and range of the impulses you've got...and there is a pretty good range out there for free.
Down side is less tweakability and more CPU (which is only a problem if your system is a little older.)

As for quality...I'm pretty picky with my reverbs, but I don't hear much in it between things like Princeton and convolution. Not enough to make me want to shell out immediately...although I probably will at some stage. I don't even see the need to buy a convolution plugin. Things like the Waves one has the advantage of excellent sets of impulses of real spaces etc thrown in for the (ridiculous) price, but the quality of SIR is so good that for my types of music, I see it as plenty good enough.

The thing that really tips it for me is the flexibility of convolution - you can use it for so many other things as well as reverb. So rather than spend several hundreds on plugins (and if I want the EMT Plate, then I also need to buy a UAD card on top) that I'd rather spend the money renting some choice gear and making my own impulses. Probably less money in the long run. Or it will be when there is the full choice of quality plugins for rooms, halls, plates, real spaces etc. And at the moment I only consider the EMT Plate and the Princeton room jobby as any competition for convolution.

To cut it short...it depends on how you work, what type of music you make, what your system is, and what you want to spend. I don't think there's a definite winner in terms of sound quality. I would be happy with the EMT and the Princeton, but I also am happy with convolution.