Reverb Subculture

[dale116dot7]

Ok, I think I understand what you're looking for here. I haven't tried anything other than a sinusoidal chorus on the AL3201 so this will be interesting. I've used 37 cycles out of 105 available so far, the AL3201 doesn't do support math except for a sinusoid or triangle chorus, so the 9S08 will have to do the modulation unless it's a sinusoid or triangle. Unfortunately, I can't read or write the chorus offset registers in the AL3201. It can also only update one instruction per wordclock plus a bit of a safety delay. Can I just fade in one tap as the other one fades out or should I glide between the taps like the LXP15 'glide delay' function? I'd guess that it should be a 'glide'. I think I can come up with random numbers pretty easily. I might have to experiment with how quickly I move taps around, too.

[Casey]

Originally Posted by dale116dot7

Can I just fade in one tap as the other one fades out?

Yes, I think this is a way to look at it. In that each tap is really two taps next to each other. The value that should be summed to the output will be a linear combination of the two neighboring taps ie.

27/32*mem[n] + 5/32*mem[n+1], then 26/32*mem[n] + 6/32*mem[n+1], then 25/32*mem[n] + 7/32*mem[n+1]...

then a memory address change occurs when the sequence reaches this point:

1/32*mem[n] + 31/32*mem[n+1], then 0/32*mem[n] + 32/32*mem[n+1], then 31/32*mem[n+1] + 1/32*mem[n+2]...

if you use 32 steps between each sample.

It's of course OK if a tap doesn't move during a given sample time, if it takes multiple sample times to set up the new move.

The lead sample will be either n+1 or n-1 depending on which direction the tap is moving.



-Casey

[seancostello]

Originally Posted by Casey

Yes, I think this is a way to look at it. In that each tap is really two taps next to each other.

How about taps not next to each other? Blesser mentions this in his book, page 269:

"Alternatively, any gain can be changed by slowly moving its value, which does not require interpolation. Typically, a randomizer gradually fades the gain of a path between two values, or reduces its gain to zero, moves the path, and then gradually increases the gain up to full value."

Also, why modulate the output taps, and not the ends of the combs? I would have thought that this would result in an overly chorused sound, but I haven't tried delay length interpolation of the output taps.

Dale: For random modulation, can you have the 9S08 change the frequency of the sine/triangle LFO to a random value + fixed offset every N samples? For the triangle LFO, this would produce a result similar to randi, the linearly interpolated low frequency noise generator found in most computer music languages since the 1960's. A sine LFO would be similar, but smoother. The random update of LFO frequency can happen at a VERY slow rate - 3 times a second has worked well for me in the past.

Nice thread. The fun little reverb project has gone from Schroeder 101 to some very cool concepts, in very easily digestible form.

Sean

[Casey]

Originally Posted by seancostello

How about taps not next to each other? Blesser mentions this in his book, page 269:

"Alternatively, any gain can be changed by slowly moving its value, which does not require interpolation. Typically, a randomizer gradually fades the gain of a path between two values, or reduces its gain to zero, moves the path, and then gradually increases the gain up to full value."

No fair reading ahead Sean!

We can go there after establishing a working reverb that can be tuned to sound pretty good.

There are lessons to be learned in working with a complete reverb and then taking the next steps.

Like you said, nice digestible steps.

[dale116dot7]

I guess what I was asking was weather Sean's approach would be ok or if interpolation (gliding) was the way to do it. Sounds like either approach would work ok. The 'easiest' way to do a gliding delay on a 3201 is to use the chorus instructions, set them to sine or triangle or something like that, and either run them or don't run them for some random amount of time. Essentially, rather than having the random number generate a tap offset, that random number will generate an integration time (at a fixed rate so the 'chorus' effect is controlled), which will be randomly run. I'll try that first. By ramping the integration rate from, say, 1Hz to 0Hz, then back up randomly over a period of a few seconds, that should result in tap movements that sort of 'move' once in a while.

I found it interesting how well adding just a couple of allpasses on the outputs would result in such an improvement.

[Casey]

Originally Posted by dale116dot7

I guess what I was asking was weather Sean's approach would be ok or if interpolation (gliding) was the way to do it. Sounds like either approach would work ok.

Gliding will work, and it sounds like you have an interesting strategy for getting it done efficiently in your system.

Sean, could you elaborate on the specifics of your approach?



-Casey

[seancostello]

Originally Posted by Casey

Gliding will work, and it sounds like you have an interesting strategy for getting it done efficiently in your system.

Sean, could you elaborate on the specifics of your approach?

The idea comes from granular synthesis, and some of the pitch shifter work I have done (the pitch shifter in SuperCollider works in a similar way):

- Crossfade between a pair of taps that are separated by many samples, up to the maximum modulation width.

- When the volume of one of the taps is at a minimum, randomly change the tap location, within a range specified by modulation width.

This also works with individual taps, each with their own volume envelope, which is closer to most granular synthesis approaches. The tap rate can be constant, or can be randomly varied each time the volume is at zero. Working with pairs of taps would help to create a more constant tap density, as you could use some form of equal power panning. You could also crossfade between more than two taps, using a multiphase oscillator.

My guess is that this technique has been used before in reverbs before (and possibly some old pitch shifters), and it seems similar to what Blesser describes in his book. The AL3201 could probably do this if the modulation is driven by the microcontroller.

Sean

[Casey]

It seems like you would hear the varying comb filter coloration of crossfading taps that are within a couple of ms of each other?



-Casey

[seancostello]

Originally Posted by Casey

It seems like you would hear the varying comb filter coloration of crossfading taps that are within a couple of ms of each other?

Yes. However, it is a time-varying comb filter. For pitch shifters, it ends up sounding way better than the fixed comb-filtering effect you get with non-intelligent pitch shifting (no looking for splice points). For reverbs, I'm not sure yet. The artifacts can sound like low bit rate MPEGs at times. I'm not using this in my current work.

I see your point, though. Perhaps there should be a minimum distance between taps that are crossfaded. In this situation, the corresponding time-invariant case would be 2 output taps from the example reverb, rather than a single tap.

Sean

[dale116dot7]

I'm going to try using the 'gated chorus oscillator' approach this morning. If you had not three, but maybe twenty taps in each loop, could you not just fade one out, then fade another in? That tap would disappear briefly but missing one of twenty for a few milliseconds might not be as noticeable as missing one of three.

[dale116dot7]

Ok, I tried this out. It does prevent 'buildup' or 'cancellation' on average. While running vocals or percussion through it, I don't really hear buildup. With a pure-sounding tone (clarinet, in this case, but the test oscillator on the console pinpoints this in an annoying way), buildup or cancellation happens but after a fraction of a second it goes away. What is noticeable when it comes to switching then staying at a particular tap is that the cancellation or reinforcement 'sticks around' then goes away. Perhaps running a very low-speed chorus in parallel with switching taps would help? ie. Rather than freezing the taps, then moving them a bit, you move them around at very low speeds, maybe a couple samples per second, then you randomize a faster move on top of that. Or is there something else you have in mind for this?

I am guessing that in 'Lexicon' terms this would be kinda-sorta like the 'Wander' parameter? Also, by using a 5ms tap movement, that works but I probably need to lower the tap movement if I lower the size of the loops (for a smaller room). Does that sound right? It would have to, I guess. Right now my 'room 'totals 416 ms, but divided down, the smallest dimension is 100ms or 33 metres. A 5ms chorus is equivalent to 1.6 metres, but if the room is scaled down in size (from 33 x 45 x 58 metres) the distance the taps move is quite a bit relative to the room size.

Since I'm doing this with the LFO's, I notice a tap movement rate where the tap movement becomes noticeable. A tap rate movement of about 100 samples per second can be noticed on pure tones or during the tail decay. Under about 50 isn't easy to notice - the more output diffusion, the less noticeable it is. In a busy song, much faster tap movements are possible unless they happen to fall in a bad time where you might hear them. When you have drum leakage into other mics, you get a funny 'whoosh' effect when a tap moves - I guess a modulation of the drum leakage, what you want to actually have reverb on, and the LFO.

I'm still curious as to how to get rid of the tape looping effect more completely, but I guess we'll get to that. The rest of the algorithm is sounding pretty good. Still some cleanup in the tap movements and further improving (actually, reducing) the tape looping should make it possible to use this in a mix. I think I'll do a mix of some kind of song in a few days so people can hear this. It'll be a mix with way too much reverb, so you can hear the processing.

So far I have 44 free instructions after adding the chorus, but I also have another 100 instructions left in the other chip. The two chips are in series (a true cascade arrangement) with only left and right audio between them. I would probably put input conditioning - predelay, input filters, maybe some allpasses on the inputs - in that part. I actually have a module written with those functions in there, but I have not put those into this algorithm yet.

[seancostello]

Originally Posted by dale116dot7

Rather than freezing the taps, then moving them a bit, you move them around at very low speeds, maybe a couple samples per second, then you randomize a faster move on top of that. Or is there something else you have in mind for this?

So, my understanding of the modulation that you are using is that it is delay time modulation, using the built-in linear interpolation, where the modulation starts and stops. Is this correct?

It seems like you would want to have constant motion in the taps, with randomly changing amount of motion, and/or direction of motion.

A tap rate movement of about 100 samples per second can be noticed on pure tones or during the tail decay. Under about 50 isn't easy to notice - the more output diffusion, the less noticeable it is.

Diffused chorusing is a much different sound than chorused diffusion. Since Casey is using some pretty big output diffusors here, low amounts of modulation would probably be smoothed out, and not sound like obvious chorusing. Once you break the LTI constraints of the reverb (in this case, it is no longer time-invariant), the ordering of the different blocks makes a difference. Some good ideas to try out.

Sean

[dale116dot7]

Yes, what I'm doing is running one of the four LFO's at any given time with the other three frozen. The times I switch the LFO's vary with a pseudo-random value. I actually read the executable code from start to finish to get that randomized value. Perhaps a better approach is to use that random value to multiply the average tap movement rate and keep all of the taps moving, though at randomized rates, which should, when integrated by the LFO, result in randomized locations. I changed the coding of the host software so the LFO sizes vary with the master loop size parameter.

I might add, scaling the loop size parameter really does make the room size change from a pretty decent-sized hall to a pretty small room. It actually sounds more like a finished reverb in the 5-10 metre region (sizing 332 m/s and the shortest loop), above that is where the looping sound happens. I've gotten it pretty close by messing with the output tap amplitudes. I am thinking that it would be better to have a loop gain multiplication between each tap, spending another six instructions over what I do already, but keeping the gains of each multiply identical, which makes the host's job trivial. The other thing is the apparent gain of one of the two output channels is quite different - the left to right level balance is a hair off. Perhaps that is intentional?

In between reverb experiments, I've been coming up with digital delay patches for that mode of operation. One thing I find with the AL3201 is not enough memory. 32k just isn't enough for really big spaces.

-Dale

[dale116dot7]

I tried distributing loop gains, and that helped the tape looping a bit, so my output tap levels are obviously not right - they need to track the RT multiplier better. But by distributing that loop gain around, the length of loop, the RT becomes a bit grainy since all gains are the same. I also tried cross-coupling the loops, but at any setting that got rid of the looping I found artifacts that I didn't much care for - strange ringing between the loops.

I tried this reverb with the distributed gains on some music, and it sounds pretty good - not quite as good as 'vocal magic' on the '91, but still, pretty good. I think the looping is the biggest difference between the quality of the sound of the '91 and this algorithm.

[Casey]

Let's save cross coupling for later. It's a great next step but it is a bit of a can of worms.

Two things to accomplish before tuning this topology.

Make sure that each loop RT and high frequency reduction is identical.

Keep the modulation moving. Just simplify the design and use a standard chorus on each tap. This is just an artifact of your implementation. There is nothing wrong with that. Work with something between 10 and 50 samples per second. And split the chorus phases exactly into 40 degrees between each tap.

If it sounds a bit to chorused, well that can be randomized better when you go to the 56k, for now just call it a feature!

What you have now is the topology of the EMT 250. If you have access to an IR set of the 250 then you have something to tune against.

I encourage you to tune your current design until you can beat the 250 IRs. To be fair, vocal and snare would be the most useful comparisons, as these are where the 250 is most often used.

The thing that will seperate this topology from the others is the snare reverb will have a tail that sounds like "cawhhhh" and not "chchshssssss". The next steps will change this, and this is a downside!

In the next step we will look at ways to further eliminate the tape loop sounds. But for now, the tuning steps I mentioned will help the tape loop problem and the left right balance. Adjust the tap gains/slope to help in both of these issues as well.



-Casey

[seancostello]

Originally Posted by Casey

This is just an artifact of your implementation. There is nothing wrong with that. Work with something between 10 and 50 samples per second. And split the chorus phases exactly into 40 degrees between each tap.

The AL3201 doesn't have the ability to arbitrarily set phases between the built-in LFOs. One option would to be to use 2 of the cosine LFOs, as these have quadrature outputs. It is easy to get 4 phases out of a single AL3201 LFO, separated by 90 degrees. Not sure what to do with that last tap.

What you have now is the topology of the EMT 250.

WOW! I had wondered if you were describing something similar, but this is cool to see. I will give this a try tomorrow. I need to program up a good low frequency random generator, but this should take a few hours at most (I have some code from a few years back that I can port over).

One question: Should there be any modulation of the comb lengths, in addition to the modulation of the comb tap outputs?

Sean

[dale116dot7]

Originally Posted by Casey

Let's save cross coupling for later. It's a great next step but it is a bit of a can of worms.

Yes, I figured that out... it's easy to make an oscillator.

Make sure that each loop RT and high frequency reduction is identical....
wrong with that. Work with something between 10 and 50 samples per second. And split the chorus phases exactly into 40 degrees between each tap.

It is a bit easier to randomize the chorus phases, it actually would be pretty easy to do. There are four LFO's with +/- SIN and COS outputs. If I randomly modulate the LFO frequency, I should get random phasing between the taps.

If it sounds a bit to chorused, well that can be randomized better when you go to the 56k, for now just call it a feature!...
What you have now is the topology of the EMT 250. If you have access to an IR set of the 250 then you have something to tune against.

I would think that setting the LFO frequencies to random would work for that. At 10-50 samples per second, with no output diffusion you can hear chorus. With the output diffusion on, the chorus basically disappears. I don't have a 250 IR, (I'm out-of-the-box) though I probably could arrange it. But I do have a PCM91, TC M5K, and a REV7 to 'tune' towards. Are there any settings on a '91 that are close? My studio doesn't quite make enough $$$ to pay for an M7 yet...

I encourage you to tune your current design until you can beat the 250 IRs. To be fair, vocal and snare would be the most useful comparisons, as these are where the 250 is most often used.
The thing that will seperate this topology from the others is the snare reverb will have a tail that sounds like "cawhhhh" and not "chchshssssss". The next steps will change this, and this is a downside!
In the next step we will look at ways to further eliminate the tape loop sounds. But for now, the tuning steps I mentioned will help the tape loop problem and the left right balance. Adjust the tap gains/slope to help in both of these issues as well.

For vocals this already is sounding quite good. I think the decay multiplier needs a bit of work - I could distribute RT gains by using a 10 to 14-bit multiply instead of 7-bit. That fixes almost all of the obvious looping (the level looping), and also I could fix the RT's between the loops to be identical. It uses a lot of cycles but I have about 25 left. Otherwise I could put in tables of loop and tap gains vs. RT. There is a hair too much grain in the RT settings and I can hear the far tail begins to loop a bit while the short one decays slightly more. If I change the feedbacks just one notch, it reverses. I think that's because I used distributed loop gain to achieve even tap gains so the effective 'bitness' of k goes down. I suspect the imbalance in levels is partly due to 4 taps on one side and 5 on the other. Another fix for this would be to offset each tap gain by a third of a bit, which would maintain the precision, but keep 7-bit math.

To me one obvious cure for looping is simply to put a loop allpass in each loop, k=0.4 to 0.5, maybe 1/3 the length of the loop or so. That should smear the 'impulse' looping at very long RT's when a sharp something happens just before silence and all you hear is the tail.

-Dale

[Casey]

Originally Posted by dale116dot7

To me one obvious cure for looping is simply to put a loop allpass in each loop, k=0.4 to 0.5, maybe 1/3 the length of the loop or so. That should smear the 'impulse' looping at very long RT's when a sharp something happens just before silence and all you hear is the tail.

Originally Posted by Casey

The thing that will seperate this topology from the others is the snare reverb will have a tail that sounds like "cawhhhh" and not "chchshssssss".

This topology is a "constant echo density" topology. Adding allspasses in the feedback loops will change this, sort of moving away from one of the charms of this very early design.

I think that getting everything possible through tuning this topology, though possibly frustrating, will establish a level of intuition as to what can and cannot be achieved with this most simple use of a couple of basic building blocks.



-Casey

[Casey]

Originally Posted by seancostello

WOW! I had wondered if you were describing something similar, but this is cool to see.

This is the basic topology, but does differ in specific details. I used the number 3 quite a bit so that the room dimension anology could be used for example.

Originally Posted by seancostello

One question: Should there be any modulation of the comb lengths, in addition to the modulation of the comb tap outputs?

Yes, this will move the frequency spikes of the feedback loops, reducing the build up and audibility of them. Good next step.

One other note, we are using linear interpolation rather loosely in this design. This will act as a moving filter which at it's deepest cut will reduce the frequencies that are roughly at 1/4 of your sample rate.

I mentioned earlier that a 2 pole input filter should be used. This is to sort of cover up this effect. If the frequency wavering becomes too audible, just reduce the input filter down from 12k to say 8k or 9k to help mask it further.



-Casey

[seancostello]

Originally Posted by Casey

This topology is a "constant echo density" topology.

"Constant Density Plate"?

[Warp69]

Originally Posted by Casey

.....................if you use 32 steps between each sample.

Would you consider a 32 steps interpolation good enough for high-end reverbs?

I currently use 256 steps - the high quality interpolators in Eventides uses 128 steps.

[dale116dot7]

Originally Posted by Casey

This topology is a "constant echo density" topology. Adding allspasses in the feedback loops will change this, sort of moving away from one of the charms of this very early design....I think that getting everything possible through tuning this topology, though possibly frustrating, will establish a level of intuition as to what can and cannot be achieved with this most simple use of a couple of basic building blocks.

Ok, I see. I've done a bit more 'fiddling' with both settings and adding and taking away 'stuff' around it. Tuning of the tap gains affects the amplitude part of the 'looping' effect. I have that more-or-less figured out. My first approach was to distribute loop gains throughout the loop, but that's inefficient. The loop gains also need to be slightly different because the loop lengths are different. I worked out a spreadsheet that should give me tap and feedback gains for a specific RT multiplier. I can just turn those into a lookup table and interpolate it.

The 12dB filter was easy to add, and yes, I also heard the filtering artifact from the chorus with the EQ turned off. But you also hear a high frequency reverb that you just don't hear in real life. Another option would be to use allpass interpolation, though I've not determined if it is possible to implement this on the AL3201.

The other thing I need to do is work on the output tap locations a bit. But with pretty close to equal spacing and the slight chorus (I set it to 0.3 Hz, which means the chorus will randomly set 5ms amplitude at between 0.02 and 0.3 Hz), it sounds pretty good to me. It shrinks with lowering room sizes.

The 'looping' that remains - and there's not that much of it any more - I think is simply the content of the input signal recirculating throughout the loop. At lower RT's (effective audible RT less than maybe two or three passes through the loop), you don't really notice it. You only notice it at the end of a song with a long RT setting, when everything else stops and you only hear reverb tail. Without increasing density and sort of 'smearing' the sound, I think you're going to hear that. Adding a few allpasses before the loop seemed to help the 'looping' dramatically, and it really smoothed out the sound in general. That would mean that in this modified algorithm there are a total of seven allpasses - two in each output, and three on the input. I can't necessarily put it in the same class as the PCM91, but I would certainly not be ashamed to use it as a vocal reverb on a mix. I find it a bit dark for snare, but I like a crisp snare sound. I should try it with more bandwidth on the input filter.

I now see how a few more of the parameters in the PCM91 relate to constants in the DSP portion, but what I'd like to know is if there is a way to get a constant density sound on it. I don't see a 'constant density' algorithm in the '91, though I might have missed it. I would guess cranking up the diffusion and cranking up definition might do that, but I don't really know how the host processor in the '91 converts a knob setting into the gains used by the Lexichip.

[Warp69]

Originally Posted by dale116dot7

I now see how a few more of the parameters in the PCM91 relate to constants in the DSP portion, but what I'd like to know is if there is a way to get a constant density sound on it. I don't see a 'constant density' algorithm in the '91, though I might have missed it. I would guess cranking up the diffusion and cranking up definition might do that, but I don't really know how the host processor in the '91 converts a knob setting into the gains used by the Lexichip.

It's only possible to get constant density with the Concert Hall algorithm by having a max value for Definition. Diffusion will control the density.

[Casey]

Originally Posted by dale116dot7

That would mean that in this modified algorithm there are a total of seven allpasses - two in each output, and three on the input.

I probably wasn't clear but I tried to mention using 3 allpasses on each output (the last being around 2.5ms.)

But it sounds like what you have is working for you, so just stick with that.

When do we get to hear it?



-Casey

[dale116dot7]

Maybe I do have three on each output, I forgot about the little ones. I don't have the code at work. But things smoothed out with more allpasses - I put them on the input.

I'll do a quick mix with 'way too much' reverb tonight, including a 'stop' so you can hear the tail. I'll try two RT settings, too, one is 'reasonable', and the other is 'RT of a canyon with a size of a high school gym. I'll also do it with and without the input allpasses.

-Dale

[Casey]

Originally Posted by dale116dot7

I'll also do it with and without the input allpasses.
-Dale

Just leave them in. It sounds like you have a nice sound now.

I think it would be more interesting to here differing amounts of chorusing using the method you came up with. A little chorusing can sound nice. I think it would be interesting to see peoples comments on how much you can "get away with" for different sources.

I would also include some mixes with a "tasteful" amount of reverb, since this is what many folks are used to judging. I assume this thread will be around for a while, and I would hate to miss this chance to showoff what you have now, before the next steps are taken.



-Casey

[dale116dot7]

All of these are a size of 33 metres (100 ms shortest delay).

New Page 1

2 = just some random settings, input diffusion set on
3 = super long RT
4 = super long RT with chorus at 3.25 Hz (default is 0.27 Hz)
5 = similar to 4, chorus adjusted
6 = no input diffusion, chorus at 1 Hz
7 = no input diffusion but a 3-tap multitap at around 0, 18, 25, 32, 45ms ahead of reverb
8 = reasonable RT and level, 'close' settings
9 = what we're starting with (dry)

You probably want to be on DSL or other highspeed, I left them as WAV so you would not hear any of the funny MP3 processing that's pretty obvious.

The input diffusion is covering up a ringing that I don't like. I think I was battling the 'loop' sound and didn't quite catch that. I have to move a tap or two in the output section, I would like to try 5 again after some more tuning of the output tap locations. A problem I have is moving an output tap on the fly and doing a chorus at the same time. It can be a bit tricky. Usually I take an output tap and just have a separate chorus delay line area, it's much easier to deal with resizing in the user interface to DSP interface.

[seancostello]

Hi Dale:

I listened to most of your settings, and it sounds pretty good to me. The main flaw that I heard is probably inherent with a 3 comb system: not enough modal density. I have been playing with a similar system today, using 4 combs (I cheated and cross-coupled them). No matter what type of modulation I use, it just doesn't sound as good as my larger reverbs that use at least twice as much delay memory.

As far as the ringing, you should see if it is from the combs, versus the output allpasses. I found that putting the allpasses on the output seems to magnify any ringing in them, versus having the allpasses on the inputs to the combs. Not sure why this is. It might make sense to tune the comb taps first, then work on the output allpasses delay lengths.

Sean

[Casey]

Wow Dale, very impressive. What is that, just one week of part time work!

I agree with Sean there does seem to be some needed tuning of those output allpasses, I'm sorry my sizes were probably not ideal.

Originally Posted by seancostello

No matter what type of modulation I use, it just doesn't sound as good as my larger reverbs that use at least twice as much delay memory.

You can increase your modal density by implementing the simplest form of the FDN. Simply take the output of loop1 and instead of feeding it back to loop1 feed it back to loop2, then loop2 to loop3 and loop3 to loop1.

Now your modal density is greatly improved and the modal peaks and valleys of your feedback loop will be much softer. This is because the amplitude of the modal peaks is proportional to the overall gain in the feedback loop. The gain will be less as the feedback path now moves through three gain reductions around the loop.

This change also tilts the design to a more balanced feed forward/feed back design as the loop now has 3 feedforward paths as well as the feed forward paths of the taps. Each of the 3 feed forward paths in the loop will provide its own modal pattern on top of the single feedback modal pattern.



-Casey

[dale116dot7]

I checked out a frequency response, and this is probably what I hear, and what the input diffusion seems to cover up quite a bit.

Below about 200 Hz, there are some peaks and valleys but generally ok. There's one buildup at 81 Hz that's a bit higher than I'd like to see. At 315 Hz there is a 20dB notch, and at about 630 Hz there's the top again. At 960 Hz, it's another notch, and again at 1545 and 2206 Hz and 2842 Hz. A classic comb. But I don't see anything that many samples apart, that's with the output allpasses turned off. Turning on the chorus changes the average so over a 20 second spectrum plot it's pretty flat. Also, without the chorus but with more allpasses on the input, it flattens that out pretty well. It seems to me that the depth of the comb response is the problem, not that there is a comb response. It's pretty dramatic. When listening to a test tone, sometimes the output signal just up and disappears, returning in a fraction of a second. But where do I look if I can't see anything that is that many samples of delay? I think it should be a 76 sample delay to generate that pattern at 48 kHz Fs.