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

[mystran]

Some of our filters got fairly noisy with less than 14 bit precision.

I thought I'd comment on this, because I would argue your statement is somewhat meaningless unless you specify a reference against which you evaluate the error.

You see, I would argue that it's not completely obvious what the "accurate" solution should be. I would love to define the ideal as the sampled response of the continuous time system after band-limitation (as required by Nyquist-Shannon), but that definition is certainly problematic in quite a few ways. By unconditional respect if you can manage this.

Now, I don't know what integrator you use (that's not really relevant once we reach the point), but BLT could be considered to have piece-wise linear input and output (and we can use the midpoints instead and get the same result in linear case). In a non-linear case one could then define the "accurate" result (ignoring aliasing) in a couple of ways: you could iterate until midpoints stabilize, or use a solver (of one type or another, probably involving more approximations) for the full step and then use trapezoidal rule (I'm not sure how to go about this, but it's still a sensible target), or you could even attempt to integrate the exact non-linear result for a linear-segment (well, maybe not, but again we could define it as the target solution).

Considering each of those could lead to a different solution, "14 bits precision" doesn't really tell that much [edit: and that's just for BLT, which isn't the only possibility]. If I had to guess what people use, I'd pick the "midpoints stabilize" as the most common approach, but is it the ideal target?

And since this is KVR, let's make it clear that I'm not writing this to dilute your credibility or anything; I'm sure you have a well-defined reference and I do fully appreciate your work in the area, but I just can't resist being my own pedantic self, and I'm genuinely interested in what the underlying assumptions about the desired solution are (because I certainly don't know what would be the best choice, yet it might have "drastic" consequences).

[Urs]

Some of our filters got fairly noisy with less than 14 bit precision.

I thought I'd comment on this, because I would argue your statement is somewhat meaningless unless you specify a reference against which you evaluate the error.

True dat.

By 14 bit accuracy we mean that within a range of -1 to +1 for possible filter outputs, the error (output - predicted_output) lies within a range of 1/16384.

This also means that we don't do Newton-Raphson, because NR can not guarantee a degree of error.

(for an exact definition I would have to call Clemens, but he's on vacation).

Urs

[Lotuzia]

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???

You don't get it ( as usual with your sliced quotes taken out of context without references ). it just seems weird to me that when talking to Xavier, who is Xils-lab, AD took an innocent rapid jab toward me ( as a Xils representative )

Imo he was not aware that Xavier was the CEO of Xils-Lab, but maybe I'm wrong.

Else if you can't see the difference between a collective entity like a company, and an individual, I can't help you. ( Though I'd like to because I really think that the guy who is not able to read properly the first 2 lines of a 5 lines test probably needs some help )

[braj]

disin·genu·ous·ness n.

Usage Note: The meaning of disingenuous has been shifting about lately, as if people were unsure of its proper meaning. Generally, it means "insincere" and often seems to be a synonym of cynical or calculating. Not surprisingly, the word is used often in political contexts, as in It is both insensitive and disingenuous for the White House to describe its aid package and the proposal to eliminate the federal payment as "tough love." This use of the word is accepted by 94 percent of the Usage Panel. Most Panelists also accept the extended meaning relating to less reproachable behavior. Fully 88 percent accept disingenuous with the meaning "playfully insincere, faux-naïf," as in the example "I don't have a clue about late Beethoven!" he said. The remark seemed disingenuous, coming from one of the world's foremost concert pianists. Sometimes disingenuous is used as a synonym for naive, as if the dis- prefix functioned as an intensive (as it does in certain words like disannul) rather than as a negative element. This usage does not find much admiration among Panelists, however. Seventy-five percent do not accept it in the phrase a disingenuous tourist who falls prey to stereotypical con artists.

[mystran]

Some of our filters got fairly noisy with less than 14 bit precision.

I thought I'd comment on this, because I would argue your statement is somewhat meaningless unless you specify a reference against which you evaluate the error.

True dat.

By 14 bit accuracy we mean that within a range of -1 to +1 for possible filter outputs, the error (output - predicted_output) lies within a range of 1/16384.

Ok, so you are doing iterative prediction of some sort. That's good enough answer, considering I wasn't entirely serious in the question department, rather it was just an observation of the point I wanted to make: the way one measures error is important.

From what little I've had time to play with different methods (I wrote my first "zero-delay" filter a week or two ago), it appears that depending on the method, one can have "noisy" solutions that are theoretically more correct, than perfectly smooth sounding solutions that on paper have significantly larger errors. I suspect it's related to the thing they call "stiffness" actually. This is interesting to me, since given a fixed CPU budget, one might have to choose between "smooth" vs. "accurate".

[eidenk]

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???

You don't get it ( as usual with your sliced quotes taken out of context without references ).

Looks like you're mistaking Kriminal for Whiterabbit...

[Urs]

I suspect it's related to the thing they call "stiffness" actually. This is interesting to me, since given a fixed CPU budget, one might have to choose between "smooth" vs. "accurate".

True. Don't worry about cpu budget. Take pride in cycles and concentrate on sound. Why else would one bother?

After all, while the most demanding plugins may be contoversially discussed, it's the sound quality that makes it worth while

[mystran]

I suspect it's related to the thing they call "stiffness" actually. This is interesting to me, since given a fixed CPU budget, one might have to choose between "smooth" vs. "accurate".

True. Don't worry about cpu budget. Take pride in cycles and concentrate on sound. Why else would one bother?

Well, I think for practically every plugin I've released, someone has complained about the CPU use. I just try my best to spend most of it on whatever methods appear to give the best bang for the buck.

[Frantz]

Don't worry about cpu budget.

Some of us don't want to buy supercomputers just to play a chord. I think there is a big opportunity for a developer to achieve great "analog" sound on a modest CPU budget.

[braj]

Don't worry about cpu budget.

Some of us don't want to buy supercomputers just to play a chord. I think there is a big opportunity for a developer to achieve great "analog" sound on a modest CPU budget.

To me it is like speakers, some may be really great at a modest price, but if you REALLY want to get to audiophile levels of precision you need to 'spend the money' in synths this seems to be system resources. I'm certain good sound can come from less, but it is a matter of good vs. great I do believe. Hapilly some developers are taking the approach to get the economy stuff out there as well, Urs just has different priorities

[aciddose]

just wanted to point out that every 24db series lossy integrator "ladder" filter i've tested fails this test. (we're talking analog, electronic here. definitely 'zero delay'.) none of them have perfectly flat feedback/hz usually because of filters in the feedback path but also because of various issues like OTAs which don't all respond with 100% accuracy across their entire range. for example the coefficient for one integrator may be different than the other three which will influence the amount of feedback required to maintain an exact peak level.

of course nothing we didn't already know. test is 100% bullshit and seems to be based upon the assumptions of an ignorant. for example the apple falls due to the unseen hand of god pushing in a single downward direction. only this no longer makes sense when the earth isn't flat or once your universe is no longer ptolemaic.

[braj]

Wait a minute... the Earth's NOT flat???

[pdxindy]

Don't worry about cpu budget.

Some of us don't want to buy supercomputers just to play a chord. I think there is a big opportunity for a developer to achieve great "analog" sound on a modest CPU budget.

There are plenty of low cpu synths for you to choose from... some of us want sound quality more than low cpu so it is good that different developers focus in different areas!

[aciddose]

if you flatten your head enough the earth will appear flat to you even if it isn't really.

http://en.wikipedia.org/wiki/Flatland

[trimph1]

It is not really flat..it is the back of a giant turtle...