New Filter request(s)

[PeterSt]

- Why are people always onto minimum phase as a means to stop pre-ringing?  I dislike the post-ringing as well, which minimum phase only makes worse (while it's messing up the phase).  What makes a solution that minimizes *all* ringing (a less steep filter) inadvisable?

If you disregard my previous answer to this question, then the answer would be : because it implies a worse THD for the higher frequencies. Or IOW, because it will violate the frequency domain. So, any reconstructing filter which is not a genuine interpolator (like AP is that) will "ring" because it integrates previous and/or upcoming samples. So, part of the energy of those other samples end up in the currently output sample which makes the current output sample actually a false one for level, but also and merely for its position in the time domain. This is a but touhg to explain in one post and without the other 1000, but actually and for result it can be seen as that the more the frequency domain is better reconstructed, the more the time domain will be violated. And the other way around : the more the time domain is respected, the more the frequency domain will be violated.

While I suggest that we can choose some middle - this is not so as far as I have seen so far. Thus, it is not so that if we let a normal (sinc) filter ring only a little, the time domain is almost still perfect. So no, this is destroyed right away and this is just because of the means these filters have to work. However, when looking closely at the outputs of the filters, there is an interesting thing to observe, and this is in a quite new direction for me also. I'll try to explain without plots, realizing that this is impossible to begin with :

When we are careless about the filter's behavior and only look at say a balance between pre and post ringing without looking at the steepness really, and THUS allow for steepness needed just thinking normally ... which would be a roll off beyond 20KHz and at 22.05 all has to be dead) then the filter will behave quite nasty. It's not only that the transients are killed (which is the unavoidable anyway as in my view), but it is not consistently behaving. It is about how the one sample sequence runs too fast into the next one which needs different treatment (lower frequency actully not subject to reconstruction) and which is because too many adjacent samples are invloved to form the current one. So, say that a filter is set to use 128 asjacent samples then it just does that regardless. But, that 128 emerges from the steepness; make that slow and maybe 12 are needed only. Filtering is as good (read : reconstruction is as good), but if Mr ol-ear Jud please can ditch some frequency in advance.

Side note : If Charlie says he likes a slow roll off, he must realize that this works to the LEFT of the frequency spectrum and not to the right. So, the response should be dead at 22.05KHz anyway, and the slower roll off means that we can not start at 20KHz, but e.g. at 16KHz. Just saying ...

Now, what I managed so far (and again I should show a plot of it) is that thinking like this (meaning : 16Khz is way than enough for me) I managed to have zero pre-ringing and only 4 post ringing samples. This is not quite true because I look at the really impacting samples as far as my imagination goes. So, there are more samples ringing but they are so minor that I can count them out (visibly). If I balance out *those* minorly ringing samples, *then* I have zero pre and 4 post. This, while I otherwise can easily have zero pre and 20 post.

The big fun (and new to me) is that I can see that such a filter indeed still kills the transients (makes sines of dirac pulses) BUT it is a totally flat filter without ripple when music is playing (which is what my special Dirac pulse train can show). So it is very well behaving.
Now, when I compare this to a normal minimum phase filter with normal roll off, it is a total mess and I envision that I see the phase behavior just in the impulse responses.

So think I already achieved something major, but I did not listen to it yet.

Peter

[manisandher]

Peter, thanks for your last three posts. I find all this stuff really fascinating. And Jud, thanks for bringing it up in the first place.

Maybe slightly off-topic in this thread, but there's something that's been on my mind for quite a while now that I felt Peter alluded to here (highlight mine):

Any processing performed in-DAC is very similar if not worse than processing done in-PC with in the middle of that the processing needed by the interface;
All needs spikey current and therewith all influence (generally let increase) jitter at least.

How does increasing the data rate from 1.4Mbps to 33.9Mbps affect the noise in the [USB] interface? And if this does indeed lead to a big increase in interface noise, is this not a strong argument for having the filtering done in-DAC?

I say all this from having had to use a number of 'standard' DACs in my office system whilst my NOS1 is with Paul (Scroobius). And for all the DACs I've used, I prefer to use no filtering whatsoever in XX. As soon as I do, the sound becomes grey and unlistenable to me. I can only explain this by extra noise in the USB line due to higher data rates.

(Incidentally, before I received my first NOS1, I used a Pacific Microsonics Model Two as my main DAC. And here, I did prefer filtering in XX before sending the signal to the DAC. But I was using a PCI or firewire interface, and not a USB interface. And also, the DAC was non-oversampling at 4fs rates, which made it prime for accepting higher data rates.)

Any thoughts???

Mani.

[PeterSt]

Hi Mani,

As soon as I do, the sound becomes grey and unlistenable to me.

I think it is clear already, but by guessing; I suppose you meant to say "as soon as I start using any XX filtering ..." correct ?

Peter

[manisandher]

Yes exactly. As soon as I use any XX filtering. Actually, not just XX filtering. I've tried with HQPlayer, and again I prefer no filtering in software. The sound is best with straight 16/44.1 going into the DAC.

Of course, I'm talking about 'standard' DACs with oversampling filters built into the chips, and not the NOS1.

Mani.

[PeterSt]

Maybe slightly off-topic in this thread, but there's something that's been on my mind for quite a while now that I felt Peter alluded to here (highlight mine):

Any processing performed in-DAC is very similar if not worse than processing done in-PC with in the middle of that the processing needed by the interface;
All needs spikey current and therewith all influence (generally let increase) jitter at least.

How does increasing the data rate from 1.4Mbps to 33.9Mbps affect the noise in the [USB] interface? And if this does indeed lead to a big increase in interface noise, is this not a strong argument for having the filtering done in-DAC?

First off, I don't think this is so off-topic because if someone perceives sound for the worse while actually it should be for the better, something is going on. But remember, it is all very hard to judge because we'd first need to be sure what the DAC does with the now higher sampling rate signal; if all is right "nothing much any more" (like your PMII example Mani), but undoubtedly not all is right (or the designer had another view, or the upsampling chip's manufacturer (from often ages ago) had another - or not much view at all ... or ...

But that not assumed, then the next subject is how robust that interface is against this noise. Or put differently somewhat : how much any means of, say, "filtering" is able to let influence less that USB noise, or how much USB cables matter and, well, endless stories.

Actually it can't be put into words what all can happen and to me it almost seems best to refer to Nick who was witness of a small part of it just by means of looking at this "interface noise" from a sort of your own NOS1 which is now at Paul's.
Let's say in advance : nothing much can be made out of it hence no real conclusions can be drawn or "what to do with what you see" except ... that you see something the most clearly. So let's say all kinds of sh*t all over at just measuring the supply to the oscillator.
Mind you, the normal solution for the NOS1 is not free of such noise either, but at least that is under control and explainable. What you see from that Dexa thing is not explainable at all because it only adds sh*t to the explainable noise, and all you can say is that it is quite worthless.

Did I drift off ?
No.

This whole clocking is just part of the interface just because that too needs "supply" and while the supply itself is (thus) more poor, all will be more poor of it.
So what I actually tried to say is that we can't apply general rules to common DACs which can easily be proven by looking at our own NOS1 and when we try to change a few things; right away nothing works any more *IF* we are allowed to look at the noise which now is in there all over and which in the end will end up at our oscillator(s). So, without measuring the effect of that on jitter, it seems clear to me that all sorts of effects will be there, and most probably not for the better (though it could by accidence).

Same with the higher rate, but which merely comes down to a faster change of bits going on and off. Bit = say 1.8V voltage needed for a bit going on, and close to 0V when it is off. So have more of that varying because of the higher data rate and you will have more noise. But per time unit more "flat" (think about one sample per minute and once per minute you will have a peak which really will be noticable).

Being on my stool of "proceed by inventing" we can say that it needs these kind of questions/answers for be to dig deeper in what's going on. I mean, it won't happen by itself and it really needs this typing and thus thinking somewhat in order to not write total rubbish (and as said earlier today, most are not much used to this, although they might think I wrote  total rubbish always anyway ). So :

When we think back about Nick's measurements on the i2s data and how the bits changing themselves already imply "music" just because of the noise it implies, this can be brought to another level if you think of this I found :
(and of which I think it is totally new)

The louder you play the more jitter.


Same story; you could say more bits change á la the i2s findings, but you can also (more easily) say the higher the output level of the D/A needed, the more voltage needed hence the more current needed ... the more noise implied. Ehm, the more jitter shows.

Again, I need plots here, and though I have the prepared I don't have access to them; they were waiting for nice times and nice topics. So later.
But it is just true ... the more loud the digital level is, the more jitter shows as varying jitter in between the peak-peak jitter which is a given (for at least the NOS1 situation).

What I did not do so far because I just never thought about it, is measuring the difference in jitter because of a different data rate towards the DAC (thus change output sampling rates from the player). But this is exactly why this "sparring" is needed and how finally all sould end up for the better again.
So straight on topic :

While for Arc Prediction the higher upsampling rate sure improves for various reasons but mainly in the reconstruction area, I am doubtful about any more normal filter regarding this, because no extra high sampling rate is needed for the reconstruction (just 2x is enough for that). The difference though is in the aliasing beyond the audio band, and when 2x is applied only, the energy will be at -40dB at 2x the sampling rate (thus 88200) plus/min the frequency of the signal. So, this is rather high, but whether harmful or audible is again that other matter. BUT, if this normal filtering means can bring down the jitter just because of less on/off rate going on in the interface (and further) ... then suddenly it is a good reason to not upsample so deep at all.

Ok, post is long enough already so I better stop for a while.
Peter

[manisandher]

The louder you play the more jitter.


Same story; you could say more bits change á la the i2s findings, but you can also (more easily) say the higher the output level of the D/A needed, the more voltage needed hence the more current needed ... the more noise implied. Ehm, the more jitter shows.

Hey Peter, are you talking about the analogue output stage here? So the more 'work' it's doing, the higher the noise it injects back into its supply, which finds its way into the digital supply? Or have I totally misunderstood what you're saying?

Mani.

[Jud]

- Once you say "bye to linear phase," is there any way to recover phase behavior before the signal hits the speakers?  I think I'm fairly sensitive to phase, so this interests me.

Yes I this is possible (whether for me is another matter) but I am not so sure this is realy needed; go look at the output of your speaker, assumed you have some filtering means in there. This is also minimum phase. And, it can already well be that what happens there can cancel out what we do in our "software filters" (in-DAC included). Can, because it would be quite conincidental.
Anyway what I wanted to say is that most of us are completely used to IIR (speaker-passive) filters and only when you are able to explicitly complain about that it would be useful to make that linear. But I'm afraid we only know the difference once the speaker puts it out linearly and we can compare.

Speakers without filtering means (wide band drivers) wouldn't show any anomaly, but now please attach your turntables (again) because otherwise (indeed !) your DAC will molest. Or use a NOS1 which outputs all unchanged, or be ahead of your DAC doing things by letting AP doing it ahead of it so the DAC doesn't do a thing any more ... if so.

Dizzy ?

Vandersteen speakers are supposedly designed to be time and phase accurate, including use of first order crossovers.  See, e.g., http://www.soundstage.com/interviews/int07.htm

[PeterSt]

Jud,

At being at the bottom of the first page I thought "OMG, another page to read !". And then I ended with page 2 and saw a third. There I gave up.

It was only at the second page where I saw that this was from 1998 which explained a few things for me. I mean, I don't think this would have survived long when brought out today. However, many things told are recognizable (for me) and at least very interesting to read.

Otherwise I don't think this is about the "linear phase" as we talk about in this topic. Or not about "linear phase" vs. "not linear phase". I can't be 100% sure but I think this is about proper time alignment only and how the phase of the different drivers won't be different compared to each other;
But that is quite a different subject.
And if vdS feels (hears) that any passive (!) 1st order filter is phase linear, then I personally think he is wrong. It will be "way more" linear compared to more order filters, but that is all.
And otherwise I don't understand. Also OK.

Regards,
Peter

[PeterSt]

Hey Peter, are you talking about the analogue output stage here? So the more 'work' it's doing, the higher the noise it injects back into its supply, which finds its way into the digital supply? Or have I totally misunderstood what you're saying?

Mani, you are inherently correct of course. But things are not as simple as that. For now (and this is the more easy explanation) think that current spikes not necessarily imply noise as such (though all could be called like that) but that it dips the supply up to eternity ... which is virtual of course. So supposed we have a digital supply and a supply for analogue, these both will still end up in the transformer. And if we have separate transformers, both still end up in the mains. And if we have a separate mains ring then this still ...
And don't underestimate this, since some people claim that power regenerators help (usually in the opposite direction when "surge" would be a subject).

Of course there's a 100 little subjects more in between this, like how capacitors should buffer which they do not and thus much more.
How 1 mile thick copper planes could help out and such.

But Mani, still :

you could say more bits change á la the i2s findings, but you can also (more easily) say the higher the output level of the D/A needed, the more voltage needed hence the more current needed ... the more noise implied. Ehm, the more jitter shows.

With the emphasis on this "say" (and my attempts to put out some English) it was my idea of making more easily clear what happens. Of course someone diving into the matters really, comes up with a remark like you did.  If I had to vote I'd be voting for the i2s merely being the culprit but with the notice that although this would be solely in the digital domain, it is not all that easy to see (for me) how the "volume" of the signal relates all the way. This doubt springs from what's really happening in there and which is "2's complement" related and where the lowest frequency determines the number of times zero is crossed and this is unrelated to the level of it. Actually, the lower the level the more crossings will take place because now the higher frequencies will do it too.

Of course I can try to make all posts to be as complicated and not understand as much as possible, so sometimes I take the easy road.

I can also say this :
I can guarantee you that the D/A conversion itself does not create the jitter I see. The why/how is another story again for later.
For myself though I now must try to reason out how the level of the signal (the due analogue signal) can incur for the jitter.
Ok, let's add some data :

A nice test signal shows less jitter than music. Peak-peak is more distinct for the test signal (like 1KHz), but the more random between the peaks is less there. Music makes the peak-peak less disdinct but fills up the area in between (read : the jitter varies more).

Time for pictures ? yes. But again not at hand and when at hand I am doing other things.

So this is "data jitter" ? well, I tried to sort that out, and to my firm belief it is not.
What would be true though is that the zero crossing would be a very discrete thing for a test signal (always exactly the same pattern). For music it is not at all.
So i2s.
And NOW it becomes mighty difficult to combine that with my being certain about the D/A is not doing that, hence how I now that; this does not combine. But after sitting back for a while, I can now think of i2s being present in two locations :
At the DAC (this was clear) but also where it is generated (in the interface).

Ok, must stop. But I think that part is set now.
So Mani, thank you for this sparring because I wasn't as far as that.

Peter (who by now forgot how ever this subject came into this topic but alas)

[charliemb]

If you disregard my previous answer to this question, then the answer would be : because it implies a worse THD for the higher frequencies. Or IOW, because it will violate the frequency domain. So, any reconstructing filter which is not a genuine interpolator (like AP is that) will "ring" because it integrates previous and/or upcoming samples. So, part of the energy of those other samples end up in the currently output sample which makes the current output sample actually a false one for level, but also and merely for its position in the time domain. This is a but touhg to explain in one post and without the other 1000, but actually and for result it can be seen as that the more the frequency domain is better reconstructed, the more the time domain will be violated. And the other way around : the more the time domain is respected, the more the frequency domain will be violated.

Ah!!  "Someone finally understands me."   

For technical readers:
Sounds a little like the Heizenburg principle (can only know position or momentum with precision, but cannot know both with precision.)

And so for those who are proponents of "44.1 carrying all of the information and reproducing the signal perfectly" like they teach in schools to students who are unable to challenge the teachers,  let's just say that if ADCs sampled 20K music at 192K, we could simply interpolate and have BOTH transients and frequency response.   We would not have to choose one over the other.  Such interpolation would also not introduce phase errors.   We'd have it all with just AP.   We'd also not have distortion after 16K.  Simple.

[charliemb]

Side note : If Charlie says he likes a slow roll off, he must realize that this works to the LEFT of the frequency spectrum and not to the right. So, the response should be dead at 22.05KHz anyway, and the slower roll off means that we can not start at 20KHz, but e.g. at 16KHz. Just saying ...

That is true, of course, for Fs of 44.1.

However, as it turns out, the slow filter being -3db at .45xFs is not of any consequence with the settings I prefer.   I prefer the SLOW rolloff filter, but upsampled 192 or 176.4.   So the -3db point is actually around 90 KHz ,  and the lowest images start 20Khz below 176 or 192.  (= inconsequential)

It it helps,  the best sound right now for me is with 4x with no filter selected (xxHE repeats samples four times).  This slow filter with these xxHE settings is unreal sound-wise.  So far best by far with OS DAC.   The filter is very open, reaching -100db not until 0.8 x Fs.   It is almost no filter at all and is probably meant for upsampling with some sort of apodizing filter.

[PeterSt]

Hi Charlie,

So the -3db point is actually around 90 KHz ,  and the lowest images start 20Khz below 176 or 192.  (= inconsequential)

It it helps,  the best sound right now for me is with 4x with no filter selected (xxHE repeats samples four times).  This slow filter with these xxHE settings is unreal sound-wise.

About the inconsequential : I don't think so. You can make the sampling rate for 20KHz material higher what you want, but when not filtered right away you will see the images and aliases all over the place. May sound nice to you (like harmonically rich) so no comments on that. But good it can not be.
Or I did not get what you were saying.

Regards,
Peter

[charliemb]

I agree.  That's correct. 

Things should be filtered.  And when filtered, there would be little from 22.05 KHz out to 158 KHz or 174 KHz.   And in which case, with 4x upsampling and a filter, the -3db around 90 KHz would be inconsequential.  And so AI or AP should provide this filtering and should sound best.

But actually, no filter selected (=repeat samples four times) sounds best right now.  And, yes, the images would be all over the place and I'm not sure where.   Have a picture of this mess?   If you do, I'd be interested in seeing it because my amps have workable bandwidth out to 100 KHz or so.

And to back up a little, I was talking about the SLOW filter on my DAC, which I think is a function of Fs, the sampling rate within the DAC.  If your new filter will filter to -3db at 20 KHz regardless of the multiplier (Fs again), 2x, 4x, etc., then we are talking about two very different filters.

[charliemb]

Now, what I managed so far (and again I should show a plot of it) is that thinking like this (meaning : 16Khz is way than enough for me) I managed to have zero pre-ringing and only 4 post ringing samples. This is not quite true because I look at the really impacting samples as far as my imagination goes. So, there are more samples ringing but they are so minor that I can count them out (visibly). If I balance out *those* minorly ringing samples, *then* I have zero pre and 4 post. This, while I otherwise can easily have zero pre and 20 post.

The big fun (and new to me) is that I can see that such a filter indeed still kills the transients (makes sines of dirac pulses) BUT it is a totally flat filter without ripple when music is playing (which is what my special Dirac pulse train can show). So it is very well behaving.
Now, when I compare this to a normal minimum phase filter with normal roll off, it is a total mess and I envision that I see the phase behavior just in the impulse responses.

So think I already achieved something major, but I did not listen to it yet.

Peter

Wow!  I also think you achieved something major.  You are the man.    How soon can I listen to this filter?  Should sound great, and very analog.

[PeterSt]

To be honest Charlie, I want to put it up for maybe 4-5 days already because it sounds so good. Don't tell !
But I have quite major stability issues with the playback itself for various reasons including unknown (think freezing PC's). But since yesterday they seem to be solved, but still it is way difficult. On a side note, the whole weekend I have been working on a non-preprocessed method which should be very fast, and that is now in a "prepared" stadium.

But otherwise envision that for external reasons at this moment I can use 16x only and thus is any file of a few minutes 1GB right away for 4 minutes or so. So I now allow 11GB for RAMDisk so I can play sort of a whole album and the RAMDisk is needed because the reading of the 24/705600 is too demanding for everything - but what everything is I need to slowly find out (hence relearn what XXHE settings (buffers etc.) will work.

Anyway I may be more eager to put up a beta of this than you are ...

Regards,
Peter