Zero Delay Feedback Filter (How to test if your synth has a )- Xils-Lab White Paper -

[AdmiralQuality]

I just watched the last two videos from this page http://www.xils-lab.com/pages/Zero-Dela ... lters.html and I'm not getting the point. You're testing one zero-delay algorithm against another zero-delay algorithm, both yours? And without a reference to a real analog or to a typical digital filter with feedback delay? (Unlike your first example video which at least compares 3 separate things.) What are those supposed to prove?

Also, the production of these clips is atrocious. Loud clicks in the editing. The filter aliasing is right off the charts. The last two spectrographs appear to be stereo, which only confuses the issue, while the first one is mono. And what's causing any difference between channels in the stereo spectrographs if all we're listening to is one oscillator and one filter?

Also the Franglais that this page is written in is nearly unintelligible.

This is not impressive at all. Nor is the addition of yet one more old wives' tale into our industry that the rest of us will now have to debunk forever. Thanks!

I just ran the test of a slow cutoff sweep with barely self-oscillating resonance over a static saw wave in Poly-Ana. I don't know if I'm getting the result that you think is appropriate or not, but I know which one sounds better to me. (Guess.)

Here's the Poly-Ana patch if anybody wants to hear it or put it up on a scope/analyzer. Just a single unmodulated saw wave going through a slow triangle modulated filter with resonance just above the self-oscillating range, but below the level that introduces clipping in the resonance (and the resultant aliasing). Oversampling is set to 16X the maximum Quality setting. But try lower Quality/oversampling settings as well, doesn't make that much difference in this case. (If I cranked the resonance right up and/or or the incoming oscillator levels were too high, it would want to alias. In which case higher oversampling would help a lot.)
http://admiralquality.com/products/Poly ... Test01.zip

And here's a Reaper project with Poly-Ana loaded with a GOOD spectrum analyzer (the one in the Xils video is crap) with parameters maxed so you can actually see each harmonic distinctly. Both Reaper and Poly-Ana run as free demos (Poly-Ana will make very short occasional silent gaps) so anyone can try this.
http://admiralquality.com/products/Poly ... roject.zip

Make sure Poly-Ana's hold button is on (just to the right of the keyboard) and play a middle C and let it run through a full sweep (60 seconds). See that little dip in the spectrum analyzer just above the resonant peak? THAT'S what the delay causes. Turn the quality down to the lowest setting (no oversampling) and the dip gets huge.

Here, I'll even take a picture...

Here's the dip with 16X oversampling.


And here it is in roughly the same place, with 4X oversampling.


See how the dip gets bigger and wider the lower the quality setting? THAT'S what 0 delay feedback should eliminate. And that's how you test for it.

You're welcome!

[Urs]

Hmmm, my main concern is that the actual benefits of delayless feedback are kind of played down here. Constant q in ideal filter models is merely a byproduct and can be achieved with all sorts of filter topologies and whatever tweaks to phase and/or gain.

The actual benefits arise in audio rate modulations, which happen with fast envelopes etc. but also in hot driven filters with multiple nonlinearities. This is IMHO where oversampling alone just doesn't help much, and where sound can go "muddy" instead of "crisp".

That said, for an anlogue emulation the amount, positioning and order of non-linear elements are as important, if not more so than instantaneous feedback, because they're directly audible.

[Ingonator]

Here's an ACTUAL, USEFUL test I used to do to see if a softsynth had a proper IIR filter instead of the FIRs that were common in many of the earlier products.

Run a low frequency square wave through the filter. Sub-sonic if you can get it. (It should sound like "tick, tick, tick, tick", instead of a tone.) Set cutoff frequency somewhere around the mid-range, and resonance to just below self-oscillation. The "tick, tick, tick" should turn into a "ding, ding, ding" as if a bell is ringing. Because the filter is, literally, ringing.

Now THAT'S an empirical test that holds up and can help verify what your ears are telling you. (Though it's been quite a while since I've seen a VA synth that doesn't pass this.)

The test you describe seems to work for my Slim Phatty, Diva, Saurus, Xils Synthix and also Waldorf Largo but not for the free Tactile Sounds Substance synth.

I have also checked with PolyAna (yep, i got it too since v1 i guess...) and found that it does not seem to pass your test at the lowest quality (the tone seems to stay at a clicking noise) but with increasing the quality setting it seems to pass it (tone seems to change to ringing bell sound).

I am not sure if this test is more foolproof than the other one as i'm not sure if Largo should have a 0df design or not.
So maybe it works for Largo because it uses high oversampling.


UPDATE:
Your pictures above seem to show the same that i found myself.


Ingo

[AdmiralQuality]

I have also checked with PolyAna (yep, i got it too since v1 i guess...) and found that it does not seem to pass your test at the lowest quality (the tone seems to stay at a clicking noise) but with increasing the quality setting it seems to pass it (tone seems to change to ringing bell sound).

You might need to tweak the resonance amount up at the lowest Quality setting to match the behavior at higher Quality settings. Poly-Ana's filter is definitely affected by the quality setting in a big way, it's not intended to be exactly 1:1 between all qualities.

Cutoff gets affected a bit too, so that will usually need to be tweaked when you change quality.

[EDIT: In fact, see the two Reaper screen caps in my post above? See how the resonance is lower at middle Quality (bottom) than at maximum (top)? Same deal, Poly's quality settings have a lot of effect on the filter behavior.]

I could probably fix that now, but at this point I feel it's more important to not make any changes to the voice architecture that will break existing patches. So, we'll call that a "feature".

By the way, that filter "ding" effect can be a powerful technique in sound design. Particularly when you've got more than one filter. (Or a whole bunch of notes playing.) But I just wanted to throw that out there as an example of an empirical test that actually works.

[xavier]

I just watched the last two videos from this page http://www.xils-lab.com/pages/Zero-Dela ... lters.html and I'm not getting the point. You're testing one zero-delay algorithm against another zero-delay algorithm, both yours? And without a reference to a real analog or to a typical digital filter with feedback delay? (Unlike your first example video which at least compares 3 separate things.) What are those supposed to prove?

Simply that there are several digital algorithms to get zero-delay behaviour, but sonically they are different. Each of them has advantages and inconvenient, all is a matter of taste

Also, the production of these clips is atrocious. Loud clicks in the editing. The filter aliasing is right off the charts. The last two spectrographs appear to be stereo, which only confuses the issue, while the first one is mono. And what's causing any difference between channels in the stereo spectrographs if all we're listening to is one oscillator and one filter?

Maybe we missed some points about stereo (we record a mono patch in a stereo track), but I don't see or hear any clips.

Also the Franglais that this page is written in is nearly unintelligible.

Sorry, I will try to learn more and proofread it soon.

This is not impressive at all. Nor is the addition of yet one more old wives' tale into our industry that the rest of us will now have to debunk forever. Thanks!

Sorry my poor English feel hard to understand what you are saying.
All the purpose of this page is to answer to the question What is a delay in digital filter, why do we want to remove it, whatever the method, Are algorithm sonically the same when they achieve this purpose.

I just ran the test of a slow cutoff sweep with barely self-oscillating resonance over a static saw wave in Poly-Ana. I don't know if I'm getting the result that you think is appropriate or not, but I know which one sounds better to me. (Guess.)

The purpose of this test was not to show how nice and good sounding the algorithms were, but pushing to their limits (fully self oscillating filter, with hot source) showing how the filter manages the harmonics.

About the spectrum analysis, I will try Reaper, good idea thanks (on windows I used wavelab, but I'm mostly working on mac and didn't find a good one after upgrading to Lion)

Best regards
Xavier

[AdmiralQuality]

About the spectrum analysis, I will try Reaper, good idea thanks (on windows I used wavelab, but I'm mostly working on mac and didn't find a good one after upgrading to Lion)

Best regards
Xavier

Reaper is a great product, I can't recommend it enough. (And it just keeps getting better and better.)

I really didn't want this to turn into a big fight. It's just that if myself and some of the other developers who've been commenting didn't respond to what we think are myths, then we're missing a chance to debunk them early on. By all means, talk on your website about what you think are the superior features of your products. We all do that. It's just that it feels like a bit of a line has been crossed when the marketing blurbs from one developer's website get presented here on KVR as if they're gospel. And it's just SO easy and obvious to debunk the method that Lotuzia is swearing by. Resonance levels could be absolutely anywhere, that alone doesn't prove anything.

I do agree, there are benefits to zero delay filter topologies. See the "dip" just above the resonant peak in my posts above. The higher I oversample, the smaller that dip gets (I believe because the resonant peak is running over it). Maybe you can try that test on your algos and let us know what you find.

[mystran]

Hmmm, my main concern is that the actual benefits of delayless feedback are kind of played down here. Constant q in ideal filter models is merely a byproduct and can be achieved with all sorts of filter topologies and whatever tweaks to phase and/or gain.

The actual benefits arise in audio rate modulations, which happen with fast envelopes etc. but also in hot driven filters with multiple nonlinearities. This is IMHO where oversampling alone just doesn't help much, and where sound can go "muddy" instead of "crisp".

That said, for an anlogue emulation the amount, positioning and order of non-linear elements are as important, if not more so than instantaneous feedback, because they're directly audible.

+3... one for each paragraph.

[xavier]

Hmmm, my main concern is that the actual benefits of delayless feedback are kind of played down here. Constant q in ideal filter models is merely a byproduct and can be achieved with all sorts of filter topologies and whatever tweaks to phase and/or gain.

As it was posted several time, this test helps to know what is involved in the filters but doesn't give an absolute answer apart for some filters (the 1 delay feedback moog ladder is the "school example")

But doing this test with more complicated filters, can show some good informations (never an analogue filter gets his resonance, quite self-oscillating at the bottom of the spectrum, dropping down in the middle, then raising at the end)

The actual benefits arise in audio rate modulations, which happen with fast envelopes etc. but also in hot driven filters with multiple nonlinearities. This is IMHO where oversampling alone just doesn't help much, and where sound can go "muddy" instead of "crisp".

In a couple of years (or centuries), this discussion won't be any more : every musicians (I hope there will still exist for a better world) will work with a 2 Mhz sample frequency. As the ears will be the same, then us, poor digital signal processing developers will have disappeared. None of the non-linearities which will be add to the signal will bring audible aliasing, filters delay will be reduced to 0.5 micro second instead of the actual 22 micro seconds. So should be unnoticeable.

In the real analogue filters, the delay exists and could be measured (0.005 micro second, less, more ?). This to say that reducing the delay to an unnoticeable value solves the problem.


This to say that oversampling a lot could solve the problem. But apart the CPU, in actual 44.1 sample frequency, it is the down-sampling which bring bad sonic characteristics.

Non-linearities and digital delay aren't part of the same problem.

That said, for an analogue emulation the amount, positioning and order of non-linear elements are as important, if not more so than instantaneous feedback, because they're directly audible.

I agree it is very important when the digital behaviour reproduces what it should
I mean, take a simple mood-ladder for instance and don't manage the feedback delay. Add any non linearities you want, the result shouldn't be good.

On the contrary, with the same filter, if the algorithm manages correctly the delay, you will get almost the same sonic result of the real thing Providing you don't reach the non-linearity parts.

So I think that managing this delay is the first step to get, for the developer, before to work on the non-linearity part.

Best regards
Xavier

[VoltAirOhm]

Sorry guys, but I only trust the "NEO-hybrid" stuff..

[Lotuzia]

....... . And it's just SO easy and obvious to debunk the method that Lotuzia is swearing by. ........

I do agree, there are benefits to zero delay filter topologies. See the "dip" just above the resonant peak in my posts above. The higher I oversample, the smaller that dip gets (I believe because the resonant peak is running over it). Maybe you can try that test on your algos and let us know what you find.

Its a Xils-lab White Paper, not mine ...

This is an open information thread, and opposite to what have been said, we have nothing to sell here, if we had taken some other audio examples than Xils Labs synths, the same people would have accused us to spoil and discriminate some "other synths", its as simple as that. I guess you can always trust imaginative people to accuse you of something. There's a proverb in France which says something like " Accuse somebody of anything you want, even if its totally false, there will be something remaining" . Heavily used by politicians, but this could as well be a rule for interwebz forums.

If some people, inbetween all the trolling, have learned something, or are now more aware of the things involved in the 0df filter debate, and can make their own opinion on if they are worth it, or more/less desirable, and/or in wich conditions, or totally useless (like you said in your earlier posts), then its all good. All opinions have the right to be expressed. (Of course if you could do it in another way than continuously giving names like liars, BS, etc etc we would certainly appreciate it, but one cant probably ask for too much )

If someone can find/propose some better audio tests to verify if a synth's filter has 0df, or less 0df, the better. Until now nobody participating in this thread seems to have something else to propose, but we would be very happy if it was the case.

If all this leads if only to a better global understanding/overall picture of all the parameters involved, nice.

At least imo something has emerged from the discussion : Like you, or Urs, or .... us, said, some other elements than 0df filters are playing their part in modeling an analog filter ( not counting interactions between elements ). Ok this one is obvious, but its important because it gives a relative importance to 0df part of the design. Establishing a hierarchy of what is the most, or the least, important thing could be the subject of another good topic, while I'm not sure a lot of people would agree on any possible hierarchy in such a topic.

[AdmiralQuality]

....... . And it's just SO easy and obvious to debunk the method that Lotuzia is swearing by. ........

I do agree, there are benefits to zero delay filter topologies. See the "dip" just above the resonant peak in my posts above. The higher I oversample, the smaller that dip gets (I believe because the resonant peak is running over it). Maybe you can try that test on your algos and let us know what you find.

Its a Xils-lab White Paper, not mine ...

Well you seem to have them in your signature. And you refer to them as "we".

If some people, inbetween all the trolling, have learned something,

...despite your efforts to mislead them...

or are now more aware of the things involved in the 0df filter debate, and can make their own opinion on if they are worth it, or more/less desirable, and/or in wich conditions, or totally useless (like you said in your earlier posts), then its all good.

No it's not all good. You are wrong. Wrong is bad.

All opinions have the right to be expressed.

Again, this is something that only people who are wrong say.

(Of course if you could do it in another way than continuously giving names like liars, BS, etc etc we would certainly appreciate it, but one cant probably ask for too much )

If you'd left it on your website where it belongs I'd never even have seen it.

If someone can find/propose some better audio tests to verify if a synth's filter has 0df, or less 0df, the better.

I just did, with an example patch, project, and pretty pictures.

Until now nobody participating in this thread seems to have something else to propose, but we would be very happy if it was the case.

I JUST DID.

If all this leads if only to a better global understanding/overall picture of all the parameters involved, nice.

At least imo something has emerged from the discussion : Like you, or Urs, or .... us, said, some other elements than 0df filters are playing their part in modeling an analog filter ( not counting interactions between elements ). Ok this one is obvious, but its important because it gives a relative importance to 0df part of the design. Establishing a hierarchy of what is the most, or the least, important thing could be the subject of another good topic, while I'm not sure a lot of people would agree on any possible hierarchy in such a topic.

I suppose it's a good topic for you guys in the sense that there's no such thing as bad publicity.

[Ingonator]

If someone can find/propose some better audio tests to verify if a synth's filter has 0df, or less 0df, the better.

I just did, with an example patch, project, and pretty pictures.

Until now nobody participating in this thread seems to have something else to propose, but we would be very happy if it was the case.

I JUST DID.
[/quote]
Agreed and i posted my own tests based on that too.


Ingo

[AdmiralQuality]

If someone can find/propose some better audio tests to verify if a synth's filter has 0df, or less 0df, the better.

I just did, with an example patch, project, and pretty pictures.

Until now nobody participating in this thread seems to have something else to propose, but we would be very happy if it was the case.

I JUST DID.

Agreed and i posted my own tests based on that too.


Ingo

Actually that was the FIR vs. IIR test.

The zero delay test is the dip just above the resonant peak. If there's no dip, it's probably zero delay (or maybe really highly oversampled).

[Lotuzia]

I suppose it's a good topic for you guys in the sense that there's no such thing as bad publicity.

No, obviously we made this topic for you

[Kriminal]

Its a Xils-lab White Paper, not mine ...

you are part of xils, and you are the one who has been pimping this highly flwaed paper from the start...now suddenly its nothing to do with you???