New Filter request(s)

[charliemb]

The only thing I'll mention here is that I've got Miska's HQPlayer working at 32/768 with the NOS1. I'm finding it really educational, to the extent that I now feel Arc Prediction may be creating some audible side effects. Offering some minimum-phase filters, and maybe also some noise-shaping schemes, as options in XX would be very welcome.
Mani.

Mani - I agree.

Peter - Not just for NOS1 owners, but also for those us whom don't have an NOS1 (like me at this time).   I've been trying HQPlayer under minimized OS thanks to xxHE and love Miska's asymFIR filter at NS9 noise shaping at 4x for 16/44.1 material.  It sounds so very analog and pure that I am often preferring that sound over even highRes.   For this to work with my OS DAC, I set the filter rolloff to slow, opposite of defaults, and makes this DAC act more like a filterless DAC.   It is very open like this, almost filterless.  My DAC is the W4S DAC2, the original.   Of course I'd also like a minimum phase fiter as well per Mani (by this I mean a min phase reconstruction filter).   

Miska's (HQPLayer) asymFIR filter rings, but the preringing is less than usual and the post ringing is more than usual and so this gives a more analog sound while sounding lush (because of the reconstruction part of this type of filter).  There is some phase shift at some point, but less so than with a minimum phase filter. 
     If I encounter any hardness from a CD, which is rare, I switch to poly-sinc-shrt-mp or poly-sinc-mp (both minimum phase).   And if hardness remains, I downshift to polynomial1 (interpolation) or minringFIR.

Please please please????

Also, can we have some noise shaping with that?   (BTW - I read elsewhere on this forum that you tried noise shaping but could not measure it.   I ask, measure it?   Please.   It is very apparent if you listen with your ears and put away the distortion meter.   And also,  forget about loosing one bit.  Apparently it is better than round-off or truncation because those end up as music dependent error, which according to some is worse than noise.)

[PeterSt]

Hey Charlie,

You have always been special with your findings and good/clear reporting. Ok, in my view that is. But maybe also there's something special going on between us. I mean how big are the odds that I start working on this on day 2555 of XXHighEnd's life (which was yesterday) while only 16 or so hours later you ask for it with a couple of "please"'s.
Of course, that post in your topic about 48->176.4 could have been a hint, but still not because you found a dozen of these kind of bugs before. Still, when I started this little project I was thinking of you ...

So I guess something of this will show up in the next version.

(BTW - I read elsewhere on this forum that you tried noise shaping but could not measure it.   I ask, measure it?   Please.   It is very apparent if you listen with your ears and put away the distortion meter.   And also,  forget about loosing one bit.  Apparently it is better than round-off or truncation because those end up as music dependent error, which according to some is worse than noise.)

And that too.
But let's say I have been ignorant till now because I don't see how any truncated bits which do so in the system noise itself can work out for the better. So please notice I measure reality (DAC output) and not theory (which can show noise at -200dB and better). And as I say so often, as long as things like Q5 etc. at 1 makes such radical differences, well, no way any dither is going to do *that*. So I'm merely in the leage of first things first, but to be honest there is no other first thing that I can see at the moment, so that is why this new project.

Regards,
Peter

PS: I think I will move this to a separate topic.

[charliemb]

Hey Charlie,
...  maybe also there's something special going on between us. I mean how big are the odds that I start working on this on day 2555 of XXHighEnd's life (which was yesterday) while only 16 or so hours later you ask for it with a couple of "please"'s.
Of course, that post in your topic about 48->176.4 could have been a hint, but still not because you found a dozen of these kind of bugs before. Still, when I started this little project I was thinking of you ...

Wow!  That's bizarre.  I think this might be the second time in about 8 months.  Maybe we have a telepathic connection. 

So I guess something of this will show up in the next version.

Thank you, thank you.  That will probably be freakishly good for us OS delta-sigma users, especially unattended.  And I predict that those who can choose a slower rolloff in our DAC settings will see bigger improvements.   And this will work in your favor because it validates the NOS1 / xxHighEnd as a concept because most the work is being done by the software filter in the PC,  and not by the filter in the OS delta-sigma DAC.  That is, your filter running under xxHE takes over and becomes dominant by far.

Regarding noise shaping,  who knows why each one sounds so different.  The only thing I can tell you is that it's audible even if it theoretically is too low in level to bother with.   Miska's NS9 sounds more airy (higher in frequency) than NS5 or Gauss1 at 4x.   Gauss1 is more evenly spread and in some ways sounds more natural.

It we can't hear the bit data influence,  then it must be the computational influence on the jitter.  Who knows.   Bottom line: it's audible. 

PS: I think I will move this to a separate topic.

I agree.  It is a different topic and also was thinking about quoting Mani and starting a new thread.  It was just easier to just hit reply and start typing.

[PeterSt]

Well Charlie, I always hate it when I can't explain something, so here's an attempt undoubtedly full with idiocy :

As you may hace read elsewhere I measured HQPlayer (mind you at 1KHz only). And, it measured worse for all the filter settings (THD I mean). Not by miles, but still. Also, all but one leave more HF beyond the audio band (22049 Hz) than Arc Prediction does. Remember, this is measurement only and I never listened to it. Now :

What if that dither or noise shaping just matters for that poorer THD to begin with; I don't recall I tried with and without dither types, but maybe that is not important. So, if dither makes the THD better, it should also be audible; All what measures better is also audibly better (to my ears), which does NOT mean that when it measures the same there are no audible differences. But this is for another many days.

So what's really relevant is whether I can improve by means of dither. And here another "watch out" : I never listened to my own attempts on this either - only measured (and saw nothing change - obviously as explained in my previous post).

And so we'll just provide the options and then we can all see for ourselves.

Something else :
I engaded that multi threading for AI the other day and had the idea it worked now. But while I never listened to it, this morning (which is 8 hours ago) I started measuring and noticed that the signal dropped after 15 or so seconds. So it took me that 8 hours to find out that this was caused by the parallel processing and things not being ready yet while some piece of the software did not take that into account. So yep, that's software, and all day I was looking for yesterday's changes which had to have caused it.

And then another thing you won't know about because it was prior to you being around :
Ever back AI worked with sneaky stuff so it could start playing while the remainder of the conversions (even of the same track) were still going on (quite normal for the other players because they convert real time). This caused so many troubles that eventually I disengaged it. But I will reactivate it, so we don't need to wait for ages before music starts. And hopefully make that decent now of course.

Regards and thanks,
Peter

[PeterSt]

Ok, I have decided to let this topic be a sort of "progress made" topic which can be useful for myself already, but which also could be informative to those interested.

First progress made :
a. The "AI" filtering means can now (Redbook) upsample to 24/705.6 so I can compare with Arc Prediction like apples and apples;
b. It sounds like sh*t.



Of course this is only a first attempt but it can't make me positive. Actually it is so bad that after two hours of listening to it I was quite dead;

The last time I really tried must have been with the predecessor of the NOS1-USB, when I developed AI to my best knowledge. Now, this is long ago and since then I can be called spoiled with so much better sound since then, but this really looks way beyond "capabilities". So with nothing much more changed than the upsampling to 24/176.4 as we know it (but thus now to 24/705.6), this is what I hear :

1. The very worse : all sounds the same, and if there is one thing I can't stand it is that.
2. I thus hear the filter (and filters do make the sound).
3. All grey mush.
4. There is now life in anything and this is because
5. All sounds totally congested. Compressed if you want. Flat. Processed.

Before it is misunderstood, please keep in mind that the NOS1 I listen through does "nothing" to the sound that I know of, nor it has been reported by anyone that it does. It is just totally neutral, and if more neutral than that can exist then I don't know of that.
But therefore it is obviously the perfect machine to observe the filters through.

Also good to know : All the processing done is through a "pre-process", so what comes from that are giant 24/705.6 files (4 minutes or so is already 1GB) and they are played by the very same means as I do otherwise. Yes, I had to increase the RAMDisk size to 12GB to be able to play a few tracks, but this doesn't matter. That all is apples and apples does.

For this standard configuration (as said, as we know it from upsampling to 24/176.4) it is also nice to observe that the more lower frequencies measure a tad worse while the higher frequencies measure better. This latter is all logic because that's what such a filter it for. However, measuring (THD) is about periodic signals (sine waves) and this by far is not music. So the transient response is nowhere and I guess this is the complete smear I hear and which makes all sound the same and grey. Btw, this is linear phase and thus pre-rings as much as it post rings, and with a fairly (or very) steep filter (roll off starts at 20Khz and is done at 22.05Khz).

Next thing I am going to do is provide a nice set of settings for a few major response changes, so I can already use them myself to easily change things and listen. But please be warned : I see no single really helping solution; the revelation was just too big after changing back to Arc Prediction.

Regards,
Peter

[charliemb]

Hi Peter

Thanks for trying.  Though I think you might be missing the point, or the points.

A brick-wall linear-phase filter (AI) is itself part of the problem, not the solution.  As such, it is not surprising to me that it doesn't sound good.

Some are saying that pre ringing is not found anywhere in nature, nor in traditional analog circuits, and therefore does not sound right to the human ear.  It shouldn't, so this line of explanation "rings true." (sorry)

That is why I'm trying HQPlayer, to try out some filters that address pre-ringing.  I can tell you that, over here, all of the filters that address or reduce pre-ringing sound more analog, more natural, less hard or harsh, less digital, sound smoother, and less distorted; and now some downsides, sound less live, less real, less detailed, move the soundstage back, sound more like a good stereo system,  blah blah blah.  Of these, the ones that ring more can sound more lush but can become riskier in that they can introduce some dirt on other tracks;  conversely,  the ones that ring less (we are talking mostly or entirely post ringing here) sound more dry / boring but are cleaner with less risk.

Some (most?) XXHighEnd users (me) are using OS DACs that only go to 192.  So why not compare at 192 or 384 with an OS DAC?  Oh, and if you do, open the DAC up as much as possible so that the DAC still filters Fs and as little below it.  This makes the OS DAC very fast and it acts more like an NOS1 will, except that it filters Fs.  (Remove the brick wall.  For this experiment, I set my DAC Rolloff to SLOW, which makes the DAC fast.)

Really, the original point was for requesting minimum phase filters and other filters that address pre ringing and AI is not it.  Over here, the best and most consistent one at 192 is not minimum phase but instead an asymmetrical filter with less pre-ringing and more post ringing.  This has full ringing but the pre ringing is shortened / addressed, and this results in a phase response that is better than that of a minimum-phase filter.  Izotope and others have these types of asymmetrical filters as options.  I encourage you to try one of these asymmetrical filters, as well as minimum phase, with both short and normal ringing.  ...and with dither (which I can confirm is usually better, but can be also be that None works best sometimes.)

I hope you don't give up!

[PeterSt]

Haha, maybe you missed some essences of my post ?
The most important one is that I didn't even start ... BUT must create an apples and apples situation.

And since today I have all running well for fast enough loading etc., so only now the job can start.

Btw asking me to listen to an OS DAC for its best sound regarding "a" filter is asking the impossible. If it sounds not on (my) par then it doesn't. So it will be all up to yourself. I'll just provide the parameters.

Best regards,
Peter

[charliemb]

Oh,  okay.  I did miss the point.  Thanks.

Things were not making sense to me because of all people you would know that a brick-wall linear-phase filter is the problem, to which the NOS1 and xxHE are a solution.  D'oh!  Sorry.   

[PeterSt]

He he, no problem !
Maybe to clarify a bit more about the history ...

At first there was only Linear Interpolation. But this was only the very eary attempt to make something out of "upsampling". So, back in 2007 (2006 for myself actually). It never worked with the notice that I first had a "normal" oversampling DAC, which later turned into a NOS DAC, but it just did not work. People (including me) tried, but never liked it. Btw, same with the Rabbit etc. means from Foobar.

Then the NOS1 was due and I promised myself to make a god filtering means for it. That became Arc Prediction; it was part of XXHighEnd ad people were structly advised not to use it (nobody had any "NOS1" anyway). But, people tried it, and automatically everybody started using it. So, something in general must be at play, and at the time I could not reason it out much.

Then the NOS1 came about and because "upsampling" now 100% sure worked for the very better, I tried to implement a normal filtering means for *that*. And well, this is actually the AI means of still today. It just did not work out but I left it in (just like Linear is still in).

Then came the next stage of people going about with software + DAC like me - Miska. And along with that all kind of Mac means like Audirvana+ and iZotope. And so, while I was not doing anything about it ... that dis not work as well. Notice though that for HQPlayer I am dependent on other's judgements because I never tried it (only measured - as said).

Now anno 2014 it is about time that I try something myself and let's say that this topic is about that.
Of course all has to be in the "XXHighEnd realm" like no real time conversions etc. because that deteriorates. This is how "AI" (now between quotes) is to be fast enough first, like your remarks about the multi threading and such. So this has been step 1 and that now is sufficiently OK. Still slow. but way better than it was. Easy calculation : with a 12 core processor it is 12 times faster for 12 tracks to load, and the time it takes is the time the longest track takes.

Next thing indeed is the post-ringing only (so bye to minimum linear phase (but maybe not)) or anything more in the middle if the post ringing becomes to much to bear.

But I have quite some more options in mind like a way more slow rolling off filter that thus has to start way more early. And thus less ringing.

So we will see, never mind I will keep on liking Arc Prediction better BUT that is for the NOS1. Meanwhile I am hoping myself to improve for the NOS1 at the same time. And it is only that I am not much positive about that. We-will-see !

Regards,
Peter

[Jud]

If I'm reading this right (may not be), you're saying filtering is hardware-dependent.  So the new filter may be good for other hardware, but you have doubts whether it can be good for the NOS1.

Can you explain a little more about this hardware dependency, if it is not just a matter of computing power?

And also, a pet topic of mine, considering minimum vs. linear phase filters:

- 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.

- 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?

[PeterSt]

Hi there Jud,

I think this will require more than one post for some answers, and while having only a bit of time now, here is a first :

What makes a solution that minimizes *all* ringing (a less steep filter) inadvisable?

I knew it would be confusing ... 

From previous posts in this topic it might now suddenly look I adhere ringing filters or something. But the anwer to your question is almost too daft to put up :
The only advice *I* personally can have is : filters should not ring at all - period. And that is Arc Prediction.

Well, you actuallly knew it (that kind of response).

So that's out.

[PeterSt]

Can you explain a little more about this hardware dependency, if it is not just a matter of computing power?

When we talk hardware, it will only be about "DAC's hardware". So of course, computer power may play a role, but this is not what I have been hinting at. Unless of course you saw that any filtering needs to be fast enough to not disturb (because it is a preprocess as done in XXHighEnd).

Otherwise ... I am not much saying that it is DAc dependent, but some people think it might be just because Arc Prediction is not sounding as the better one.
But I never judged that and all I can say is that A.P. was not made for any other DAC.

Ok, this is all too fast typing. Must go now ...
More later !
Peter

[Jud]

Hey Peter, I don't want to interrupt any thinking you have on these questions, because you're off to a very interesting start.  I only wanted to give a brief mention that when I referred to "computing" power, I didn't mean computer power.  Instead, I was thinking of something involving the DAC chip(s) or FPGA or whatever in the DAC, and wondering if that was the reason you were saying AI might be better for other DACs but not the NOS1, where Arc Prediction was superior.  Or perhaps it is the delta-sigma modulation in other DACs?  Just part of my question about hardware dependency, specifically referring to the DAC.

[PeterSt]

To others : Don't let yourself overwhelm by out of context technicalities with the explicit notice that I hardly ever talk(ed) this kind of way in my own forum. But elswhere I am more or less able to do that, and Jud knows it.

Instead, I was thinking of something involving the DAC chip(s) or FPGA or whatever in the DAC, and wondering if that was the reason you were saying AI might be better for other DACs but not the NOS1, where Arc Prediction was superior.

Hi Jud,

No, that can not be my reason because the NOS1 is just that and it can not contain any processing for filtering (or it should be done in the current FPGA which is not very likely). However, we could see the subject a kind of upside down :

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.

If we indeed skip a 1000 other posts about this (me involved one way or the other) than by now it has gotten so much more complex that we better stop attempting to reason out what results will really be, but let's just have one example :

First of all we must acknowledge that something like Q3,4,5=1 of 1.186 makes a hole world of difference for SQ. So, those who can't agree with that (or don't perceive any difference from XXHighEnd's settings in general) better stop reading. But the Q3,4,5 is just a best example of how from a very long distance hence very indirectly the performance of the DAC is influenced. So for those who agree (and understand the concept of this to some extend) we continue ...

If this very indirect means already does its intended job, what about the far more direct and heavily impacting processes in the DAC itself ? You can call that devistating. So, a bit outside the subject, this is how the NOS1 does totally nothing or at least nothing which could be avoided. But still a lot is going on (it has to be) but anyway this is audible (that it does nothing).
Again on a side note, this is also how the heavy processing of AI has to be a pre-process, because since anything what we do in the PC matters, this should be out of the way once we press Play.
Thus, all consistent thinking and consistent behavior for the results of it (if you'd ask me of course).

Still there ?
Now, when Charlie says (and keep in mind it is just an example of how complicated things had grown) that he likes a slow roll off better, he will already be correct (on audible results) because that requires less processing in his DAC. So, a slower roll off means less order filters and any "order" can be seen as an additional processing step (which is already more than linear - thus progressively more). And thus simple : imply less processing and your DAC sounds better.

Well ... nobody can be really sure about that because what we do is influence jitter and nobody really says that any kind of more processing implies more jitter; it can easily be the other way around because the more processing the more "even" noise will be implied (depending on the speed / frequency of the current peaks) and the more even the noise the more "random" the implied jitter will be.
Fact is (for me) : The way the filter works influences jitter and since this is of vast importance, choosing another filter already changes SQ for that reason alone.

Fun ? maybe not, because this is totally out of control.

Meanwhile of course, the changed (thus other) filter heavily impacts just the same. This too makes it not so funny because apparently we *have* to change two variables at the same time while functionally we only (want to) change one. It just is so.

Not very much related but still mentioning it because it could clarify better my thinking :
Still thinking upside down somewhat - if Arc Prediction would be implemented in-DAC now *that* coincidentally would mean the leanest way possible of applying a filter. So/and, people may not realize it, but A.P. surely *is* a real time filter, though in PC software. And how much CPU usage you see ? well, around zero. When moved to the DAC it would mean the lowest current spikes (lowest in Amperage) and its spread would also be very even (which any normal filter can not be ("even")). So lowest jitter and most random like (or white like if you want).

When the latter is reflected back on daily practice, you see that this lowest means of noise implication is now also at the distance (in-PC) and theoretically still less immpacting than when it would happen in-DAC. But :
What also follows from all is that when this would be a pre-process (so all arranged for before playback starts) it's even less noise implying and using less CPU. Maybe.

Maybe, because - and this is the interesting part of it for me - ... because now the files are huuuuge (all 24/705.6) and now all the processing goes into the reading of the files which has to be 16 times higher for the higher sample rate alone, but ... For AI this *has* to happen, or better, happens by now. So, if I can do it for that, I can do it for AP as well. Sort of. Anyway it would be an interesting comparison if I'd allow AP to be a preprocess as well (for choice).
(but I don't think I will do that because it's a sort of gadget and can't have any priority really)

Maybe more in a next post.
Peter

[PeterSt]

- 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 ?