Reverb Subculture

[thermos]

Originally Posted by Casey

David spoke with me many times and in great detail about the concert hall algorithm and the modulation methods it used. PM me.

AFAIK David does still have the 224 code base. It was the 480 code base that was somehow destroyed.

In fact, David has wondered as to whether it would make sense to re-release the 224 as a plugin. Perhaps he and Barry will get together. Barry has mentioned to me that he would like to re-release the 250. We were speculating on another thread that perhaps this is the reason that UAD has been talking up the 250 in their recent news letter.



-Casey

What stops David and Barry from making plugs of them? Is it really whether or not people would be interested? Thats got to be one of the silliest questions and most obvious answers in the history of mankind. I hope we see those plugs some day. That would be a huge contribution to society.

[seancostello]

Originally Posted by Casey

David spoke with me many times and in great detail about the concert hall algorithm and the modulation methods it used. PM me.

AFAIK David does still have the 224 code base. It was the 480 code base that was somehow destroyed.

In fact, David has wondered as to whether it would make sense to re-release the 224 as a plugin. Perhaps he and Barry will get together.

I love this thread.

[Casey]

Originally Posted by thermos

What stops David and Barry from making plugs of them? Is it really whether or not people would be interested? Thats got to be one of the silliest questions and most obvious answers in the history of mankind. I hope we see those plugs some day. That would be a huge contribution to society.

I agree.



-Casey

[dale116dot7]

Originally Posted by Casey

AFAIK David does still have the 224 code base. It was the 480 code base that was somehow destroyed.

The code is obviously contained in many hundreds of EPROM's out there. If you really want source code regenerated from a binary image extracted from EPROM's, that technology exists. It's painstaking work, but I've had to do it a few times when source code went missing (ie the computer was stolen) or when a subcontractor went out of business and disappeared without leaving source code. But it can be done. If you wrote it the first time, and had to decipher it from an EPROM, it's not that bad. If it's written originally in assembly code, it's easier than if it was compiled from some other language. Not bad if you wrote the code in 'C' but a bear if someone else wrote the code and you are the one having to reverse it. If it's all fixed point (which my engine control projects always are), it's a lot easier than floating point (which I never use anyways). If you own the compiler that was used, it's also a lot easier because you can reverse the libraries and that gets all of the math subroutines defined. Try tracing through a floating-pont divide routine... or is it a multiply routine or a square root.. I just can't tell.

[Casey]

I need to correct myself.

As I was thinking back over the related conversation, it wasn't that David was thinking about doing a plugin, he was wondering about doing new algorithms for existing 224s.

Sorry for confusing the issue.



-Casey

[zmix]

Originally Posted by Nobody Special

Hmm. What's a programmer supposed to sound like on a B.B. thread?

zmix, I'm sorry you've taken such a dislike to the PCM96. There are an awful lot of very good people in the industry who've told us it's the best-sounding box we've ever made. I do a bit of classical recording on the side, and I think it's head and shoulders above our earlier boxes for that application.

Since many of the complaints seem to have something to do with modulation, let's talk about it. The purpose of modulation in Lex boxes has never been about being obvious--it's always been about breaking up room modes and improving the spectral performance. In the case of the old ConcertHall algorithm, and to a lesser extent the RandomHall algorithm, this had mixed results. While spectral performance was improved, the modulation was obvious. For a great many users, this turned out to be a plus. They loved the effect. David Griesinger, who wrote those algorithms, was never wild about that effect--especially in ConcertHall. Getting any information from him on exactly how the modulation worked was like pulling teeth. He simply wasn't interested. That's why the 96 sounds a bit different from the older boxes. Unfortunately, the original 224 and PCM70 code was lost long, long ago.

The newer algorithms like Hall and Room have a much more successful modulation, if you consider the original goal. Unless pushed to an extreme (spin anywhere past about 2, 2.5 Hz), you can't really tell it's going on.

So as I said before, sorry you don't like it. But I'll still pop in whenever I can and answer questions and comments.

N.S.

Dear N.S.

In posting my comments and various examples I intended not to simply complain, but to open up a discussion with other reverb aficionados and to share what I am hearing in these algos, what I am missing, and hopefully to inspire Lexicon for further development of the PCM96 platform.

I have been mixing records professionally since the early 1980s. Back then I used the Lexicon 224, 224x, and later, the PCM 70. I sort of "lost interest" in new Lexicons reverbs around the time the 300 / 480 were introduced.

The primary attraction to the PCM96 for me was the claim that the original "Concert Hall" algorhythm had been reintroduced. I bit hard on that lure. Imagine a 224 with instant recall! When I received my PCM96 I immediately began recalling some of my favorite 224 settings (by ear, not copying parameter values... which were obviously different in the PCM96). I could not get it happening and instantly buyers remorse set in. I felt cheated. OK, I calmed down and went through every algorhythm trying to program some useful starting points for future mixes.

Each algo shared a similar characteristic that I had never heard in my favorite Lexicons of the past. I can pinpoint the characteristic of the PCM96 as having excessive ringing in the lower frequency modes and an overly "tonal" quality to the reverb, certain notes would simply linger unnaturally and the end result was a muddy track with altered chord voicings. Using the modulation to break up these modes lead to a lumpy response and odd image shifts. I found that my programming became focused on taming artifacts rather than pursuing the sound I heard in my head. I have an associate who offered to send me his 224 as a reference, which should prove interesting. In the meantime I hooked up a PCM-70 and started tweaking the concert hall algo. I felt completely satisfied by this version of the algo. I went back to the PCM96 and again it sounded dark, ringy and the tail just sounded weird and out of phase (polarity). I experimented with the 'tail width' parameter and discovered a bug: The parameter is offset by 45˚... at 45˚ the output is mono (and the display reads "narrow stereo"), at 90˚ the output is stereo (the display reads "mono"), and at 135˚ the tail is mono with the L/R in inverse polarity. At no setting could I match the width or depth of the PCM-70... Also the modulation in the PCM96 is unable to produce the subtle magic of the 224 / 70 it always sounds obvious and cyclical.

I can appreciate that the original 224, 224x and PCM-70 code was lost "long, long ago", but 224s and PCM 70s are readily available and a serious effort to discern the operative mechanisms by listening seems not to have taken place. I have managed to create reverb algorhythms which do produce the sort of subtle yet musical modulation, as have many others here. The PCM96 version of the Concert Hall seems to be modulating the wrong places in the algo. Also, in the description of the PCM96 algos here on Gearslutz the "Concert Hall" description reads like a data sheet on an obsolete electronic part (not recommended for new designs) and the "Random Hall" is described as being "...more responsible for the term "Lexicon Sound" than any other reverb." I say that this is an outrageous statement. Lexicon's reputation and subsequent fame was built on the back of the venerable "Concert Hall" algo (and the CD plate). If it were not for the ubiquity of that algo in the early 1980s there simply would not have been a 480 etc. You would think it deserves a bit more care than It got in the PCM96...


Not happy to simply opine, I decided to measure the differences in the basic modulation LFO. First of all the PCM-70 modulation oscillator is a mixture of a 2.3hz triangle wave and a much slower random waveform (around 0.5hz). The chorusing LFO in the PCM96 has a maximum rate of 1.5hz, and is not randomized.

I am delighted that Casey has offered to discuss the 224 Concert Hall modulation with you. I think the PCM96 as a platform has tremendous potential and I'd love to see that realized.

I hope that Lexicon take these efforts to illustrate and educate seriously and don't rest on the "Other People like it" argument.

-Chuck Zwicky

[Froombosch]

Originally Posted by zmix

Note the low density, high modulation and late 1970s primitivism....

This sounds a lot like an effect and not a reverb, but I can imagine where it can be used. It has alot of Eventide style inside.

I'll look if I can add H8000 / Eclipse reverbs to this thread...

[Froombosch]

Originally Posted by Casey

I agree.



-Casey

+1

[Froombosch]

Originally Posted by Casey

I need to correct myself.

As I was thinking back over the related conversation, it wasn't that David was thinking about doing a plugin, he was wondering about doing new algorithms for existing 224s.

Sorry for confusing the issue.



-Casey

Converting the 224 algo to a new machine would be an awsome plan. Or placing it inside for example the 96 platform. An interesting job for a good dsp-engineer...
Not all vintage stuff is better, but these old machines have a really high quality level.

[zmix]

Hi Harrie,

I agree. How nice would it be to have those early algos preserved in a representative form! Look at how retro-obsessed the plugin market is at the moment.

-CZ

Originally Posted by Froombosch

Converting the 224 algo to a new machine would be an awsome plan. Or placing it inside for example the 96 platform. An interesting job for a good dsp-engineer...
Not all vintage stuff is better, but these old machines have a really high quality level.

[Froombosch]

When I watch GS I see that there are quit a few oldies high on the 'best reverbs' list. That will be partly because of the hype, partly because of the sound. There are a lot of great sounding old machines to be ported over to newer formats.

These oldschool algos are fun to play around with. Generating enough randomness is important to making smooth and lush reverbs with this kind of reverbs. It is awsome to see ideas on this topic in this open forum.

When building reverbtanks (three series of four small delay blocks) I listen to first series of delays, spacing them up, then adding the next series. Then tweaking the second delaytimes so it generates the most smooth (least ringy) responce. I also did this with the third series. The outcome is a bit different when comparing it to the golden ratios known.

I am using VSIG on the H8000 for developing new algos. ATM we are porting some of those algos over to VST-format. I am amazed how much difference these ports make. VSIG is fast and fun to work with, but not always superaccurate.

[Casey]

Let's talk about coloration in reverb caused by allpass filters.

The first thing that any designer learns is that allpass filters are great at increasing echo density when placed in the inner loop of a reverb algorithm. This is often desirable and the reason for using the allpass filter in the first place. Unfortunatly, the next thing he learns is that while an allpass is flat in the frequency spectrum, it has a comb shaped group delay characteristic that causes ringing in the tail because certain freqencies are delayed longer than others. Since the group delay of an allpass filter is comb shaped, it is regular and leads to a prounounced metallic sound in the tail.

OK, I was wrong, the first thing that a designer learns is not to put allpass filters in parallel. Every text says so, right? Why is this? Well since the group delay of an allpass is comb shaped, two differing allpass filters in parallel will cause outputs that have phase relationships that are directly related to the super position of the two group delay comb shapes.

And yet, we see many designs that have allpass filters in parallel. No? Even if the inner loop consists of allpass filters in series, look again. Follow every input path to every output path and often you will find that the inner loop does have many paths that place the allpass filters in parallel.

So how do these algorithms get away with breaking this fundamental rule of no allpass filters in parallel? Well, the best answer is to trace every input/output path that places the allpass filters in parallel, and make sure they are spaced out enough (say > 200 msecs) that the sound in each path is decorrelated just by being far enough away in time. Much closer than this and coloration starts to show up in the lower frequencies.

If this is not possible, (and it rarely is) then randomizing the allpass filters is typically used to keep moving the comb shaped group delays so that the ear doesn't pick up a constant coloration.

A method has recently been published by Blesser, which modifies the allpass filter in order to eliminate the ringing tail problem brought about by the comb shaped group delay. At first this seems like a pure win, but in reality there is a price to pay in the coloration of the running reverb. Blesser's method alters the allpass filter coefficients so as to turn the allpass into a comb filter. Essentially adding coloration to the running reverb in order to eliminate metallic sounding tails. This is particularly problematic in that the coloration is constant.

It is also beyond the abilities of typical computing resources to randomize these modified allpass filters given the math outlined in the Blesser publication. So nothing can be done to eliminate the primary coloration caused by parallel (in this case modified) allpass filters.

For years we have been tought to listen for impurities in the reverb tail. It is just as important to listen for unnatural coloration in the running reverb.



-Casey

[zmix]

Awesome and thought provoking post, Casey...

Parallel Comb filters are not all that unusual in some reverb topologies, are they? In the Lexicon 224x there were some "constant density" plate algos, were these based on such a topology?

Also, someone at Lexicon mentioned that the PCM96 hall employs a new (patented) technique developed by Barry Blesser, do you suppose the paper you mentioned has anything to do with the new Hall algo?

[Froombosch]

Why not delete these all-pass filters in the tank and only set them in series before the reverb tank?

[seancostello]

Originally Posted by Froombosch

Why not delete these all-pass filters in the tank and only set them in series before the reverb tank?

Allpass filters in a tank result in an echo density that builds with time, as found in real 3-dimensional acoustic spaces, as well as 2-dimensional spaces such as plates. Allpass filters in series before the reverb tank will result in an increased initial echo density, but no build of density as the reverb decays.

Sean

[zmix]

Originally Posted by seancostello

Allpass filters in a tank result in an echo density that builds with time, as found in real 3-dimensional acoustic spaces, as well as 2-dimensional spaces such as plates. Allpass filters in series before the reverb tank will result in an increased initial echo density, but no build of density as the reverb decays.

Sean

Harrie, Sean,

Isn't this topology used in Jon Dattorro's algo?

[dale116dot7]

Originally Posted by Casey

So how do these algorithms get away with breaking this fundamental rule of no allpass filters in parallel? Well, the best answer is to trace every input/output path that places the allpass filters in parallel, and make sure they are spaced out enough (say > 200 msecs) that the sound in each path is decorrelated just by being far enough away in time. Much closer than this and coloration starts to show up in the lower frequencies.

If this is not possible, (and it rarely is) then randomizing the allpass filters is typically used to keep moving the comb shaped group delays so that the ear doesn't pick up a constant coloration.

But if I look at a loop algorithm in a certain well-known reverb, I see a lot of allpasses that are closer than 200ms. In fact, all of them at most sizes less than a canyon. Somewhere around 13ms at the smallest size setting. But the allpasses on either side of that delay are 8.9ms and 4.7ms - the total of which is about the same as the delay between them. Is it more the ratio of allpass length to delay length that is important here? Or to keep the amount of delay within allpasses of the loop to less than, say, 1/3 of the straight delay time in the loop?

Maybe modest gains and enough delay time as a ratio of the allpass time would prevent the effective paralleling of combs - ie. a delay of 13ms at the end of a 4.7ms allpass would allow just under three passes through the first allpass before hitting the next allpass?

Then what about the 'black hole' reverb (I think it was on an Eventide 8000) which is simply a sequence of something like 32 allpasses in series? It has an interesting sound - a very slow buildup and a decay that is about as fast as the buildup. Very unnatural sounding to me. Almost like a tunnel drive-by except without the doppler effect.

[Froombosch]

Originally Posted by seancostello

Allpass filters in a tank result in an echo density that builds with time, as found in real 3-dimensional acoustic spaces, as well as 2-dimensional spaces such as plates. Allpass filters in series before the reverb tank will result in an increased initial echo density, but no build of density as the reverb decays.

Sean

Correct. On the other side we are using modulation to decorrelate the reflections. This has no equivalent in normal hall reflections. Maybe moving walls

[Froombosch]

Originally Posted by zmix

Harrie, Sean,

Isn't this topology used in Jon Dattorro's algo?

I do not think so. The diffusers divide the reflections more after every feedback return. So this topology had increased amount of reflections.

[Froombosch]

Originally Posted by dale116dot7

Then what about the 'black hole' reverb (I think it was on an Eventide 8000) which is simply a sequence of something like 32 allpasses in series? It has an interesting sound - a very slow buildup and a decay that is about as fast as the buildup. Very unnatural sounding to me. Almost like a tunnel drive-by except without the doppler effect.

I do not know how they made it. I can try to figure out using VSIG. I love the sound of it and use it as an effect on a lot of my recordings.

[dale116dot7]

Originally Posted by Froombosch

I do not know how they made it. I can try to figure out using VSIG. I love the sound of it and use it as an effect on a lot of my recordings.

I looked at it in VSIG and I understand how it's put together but it sort of breaks the allpass separation guideline.

[seancostello]

Originally Posted by dale116dot7

I looked at it in VSIG and I understand how it's put together but it sort of breaks the allpass separation guideline.

Each output channel of the Black Hole reverb consists of a single series chain of allpasses, with the output taken from the end of the chain. If you look at the output channels in parallel, there are parallel allpasses, but the issues that Casey is talking about are less objectionable when the signals are being mixed in the air, as opposed to in the box.

The structure uses a single chain of a bunch of allpass delays for a mono input signal (derived from a sum of left and right input signals), and then has 2 parallel allpass delay chains to decorrelate left and right signals.

[zmix]

Originally Posted by Froombosch

I do not think so. The diffusers divide the reflections more after every feedback return. So this topology had increased amount of reflections.

The Block diagram and parameter list for the Ensoniq DP/Pro implementation show six diffusors in series feeding the reverb tank. I'll listen to some impulses to see how the density builds up....

[seancostello]

Originally Posted by Froombosch

Correct. On the other side we are using modulation to decorrelate the reflections. This has no equivalent in normal hall reflections. Maybe moving walls

Variable speed of sound, due to temperature variations within a large space. A small amount of temperature variation in different places within a hall, or moving air currents, will cause very small random pitch changes of the reflections. Multiply this by a few orders of reflections, and you can have a fairly significant spread of frequencies over time.

[zmix]

Originally Posted by seancostello

Variable speed of sound, due to temperature variations within a large space. A small amount of temperature variation in different places within a hall, or moving air currents, will cause very small random pitch changes of the reflections. Multiply this by a few orders of reflections, and you can have a fairly significant spread of frequencies over time.

According to Dr. Blesser's citation in his "artificial Reverberation" paper , the speed of sound varies due to thermal currents at an (average) equivalence to 2ms at a rate of 0.2hz.

If you want to simulate this effect you need to tailor your randomization to resemble wind! The largest DSP hogs in my reverb algos are the randomizers....

[seancostello]

Originally Posted by zmix

Harrie, Sean,

Isn't this topology used in Jon Dattorro's algo?

Yes it is. The Dattorro algorithm, at least as published in the JAES, uses series allpasses both within a feedback loop (to build echo density over time and decorrelate the recirculating tank signal), and outside of the feedback loop (to turn impulsive sounds into tight clusters of echos).

[seancostello]

The Black Hole reverb points out one of the fundamental issues with allpass delays in series, or for that matter any sort of filtering operation in series: As the order of the series increases, the impulse response of the output begins to resemble a Gaussian bell curve more and more. This causes a very distinctive "fade in" sound that can be used to great effect if intended, as in the Black Hole reverb, but will cause problems if not intended.

I have heard simpler digital reverbs, such as the SilverVerb in Logic, that are clearly based around comb filters with a lot of series allpasses embedded in the delay loop, and a single output tap at the end. The output of such reverbs can be very "open," but will also demonstrate that fading in quality, as opposed to an impulsive attack followed by an exponential decay.

The solution is to tap in between allpass stages, as shown in the Gardner and Dattorro papers, and later on in the Dahl/Jot papers with the absorbant allpasses.* This allows you to grab the signal before it becomes overly "Gaussian." Of course, this also creates the parallel allpass artifacts which Casey refers to above. So there is no neat and tidy solution, just a bunch of compromises and heuristic tricks.

Or, you can avoid using allpasses altogether. Lots of reverbs only use short allpasses outside of the feedback loop (if at all), and rely on a combination of parallel combs, feedback delay networks, and/or cleverly arranged output taps to get the required echo density. Having said that, so many of the classic digital reverbs seemed to use the allpasses embedded inside of recursive loops, that the artifacts are probably considered desirable for many listeners out there.

I was starting to write about this stuff on my blog (The Halls of Valhalla), but it seems like there are <10 people out there that are really actively pursuing this stuff right now, and most of them are posting to this thread.

Sean

* "Absorbant allpasses" are subject to a patent by Creative Labs. However, the use of them embedded within feedback delay network structures (which can include a single comb) was first mentioned in Jot's 1992 PhD thesis, so the validity of the claims concerning those allpasses is questionable. Plus, Gerzon showed frequency dependent allpasses in his 1972 and 1976 papers, and Date/Tozuka showed analog circuits for implementing these as early as 1966.

[Casey]

Originally Posted by zmix

a new (patented) technique developed by Barry Blesser

I should have been more clear. When I said "Blesser published..." I was referring to his patent.

The patent decribes two different ideas. The idea that I was discussing is what Blesser refers to as the "notchpass filter", which is essentially a modified allpass filter, with all of the attendent issues associated with using allpass filters except for the "metallic sounding tail issue".

Eliminating this issue is a reason to use this method, I was just pointing out the downside issues that a programmer should consider before replacing standard randomized allpass filters with the Blesser modified allpass filter.



-Casey

[zmix]

Originally Posted by Casey

I should have been more clear. When I said "Blesser published..." I was referring to his patent.

The patent decribes two different ideas. The idea that I was discussing is what Blesser refers to as the "notchpass filter", which is essentially a modified allpass filter, with all of the attendent issues associated with using allpass filters except for the "metallic sounding tail issue".

Eliminating this issue is a reason to use this method, I was just pointing out the downside issues that a programmer should consider before replacing standard randomized allpass filters with the Blesser modified allpass filter.



-Casey

I've read that patent. I believe that this 'notchpass' is incorporated in the new PCM96"Hall"...

Let's ask Nobody Special...

[Michael Carnes]

Originally Posted by zmix

I've read that patent. I believe that this 'notchpass' is incorporated in the new PCM96"Hall"...
Let's ask Nobody Special...

Yep. It's in the Hall, the Chamber, and the reverb tail portion of the new Room algorithm. There are of course other new things in those algorithms related to frequency management, but Barry's idea made some noticeable improvements.

N.S.



Since you've been talking about allpasses, I'll pass along an observation. Although the traditional view is that allpasses increase reverb density, that's only true up to a point. The number of reflections per second is still limited by the Nyquist theorum (and probably by the auditory cortex as well). With well-chosen allpass values, that limit can be reached quite rapidly. After that point, the significant effects are phase and frequency.