Interesting DAC Design exchange on that other site!

[cfmsp]

 

http://www.audioasylum.com/forums/pcaudio/messages/6/67479.html

 

begging Chris' forgiveness, I thought this thread on AA was an incredibly interesting conversation between two very knowledgable posters - John Swenson and Todd Kreiger.

 

For those interested in understanding some of the issue impacting current SOTA DAC design, this is highly recommended!

 

 

I was particularly interested in the comments on RFI, which are only a small portion of their conversation.

 

JS:

"The net result of all this is that it if you are willing to go with high sampling rate AND get a really good high sample rate implementation that doesn't muck something up somewhere along the way you CAN get sound that does what you want, that can sound as good as vinyl."

 

TK:

If the RFI problem could be addressed, I'd agree with you, but higher rates mean higher RFI...... The ideal oversampling rate IMO is 4x !!!

 

JS:

I agree with you my favorite oversampling ratio is 4X.

[...]

You bring up RFI, there are several types that can be a problem. One is RF riding along with the analog signal on the output of the DAC. This is actually fairly pervasive in the industry and is not all that easy to get rid of. I go to great lengths to minimize it in my DACs. Its NOT as easy as just slapping an RF filter on the output. The whole grounding scheme of the DAC has to be carefully engineered to get this right.

 

Another is radiated fields, this should be well under control, but surprisingly is not.

 

The other is RFI sent back into the power cord and on to other components. Some components do a very good job of this and others do not.

 

 

 

enjoy,

clay

 

 

[icebreaker]

That one post by TK has answered many questions I have asked in the past but previously never received a convincing reply. Especially about NOS and oversampling, imaging, filters...

 

Its somewhat reassuring that TK (at least) has confirmed some of my semi-informed suspicions in regard to trade offs in DAC design!

 

Now back to reading the lengthy thread.

 

Thanks,

 

[cfmsp]

Yeah, I thought John / Todd had a clearer view on some of the issues that I don't yet understand to my satisfaction.

 

Perhaps, Peter, I_S, CG or others will chime in.

 

clay

 

 

[PeterSt]

Having read only the post you linked to, all I can add is this :

 

1.

For NOS, upsampling is a means of filtering by itself (it removes the HD coming from the too rough digital stepping from the audio band).

 

2.

The way to upsample is key to the sound. Also, here the confusion starts (with my idea of John operating more empirically than scientifically -> bad for understanding but equal to my own means of operating in this area :~)

Normally (the filtering chips referred to) the upsampling is *needed* to let the filter operate, hence the filter achieves its filtering by the means of upsampling. This is based upon (and measured against) the reconstruction of individual waves, which in the long term (read : many subsequent cycles of the same frequency) indeed reconstruct the original wave. Sadly practice is not about subsequently same waves/frequencies, and this is where normal filtering flaws.

 

3.

I too created my own filtering (Arc Prediction) which also needs upsampling in order to work (and where 4x is sufficient, 2x is not). However, Arc Prediction doesn't depend (or calculates against) many cycles, but can do with a single (half) wave. This means no phase distortion and no ringing (pre or post).

 

4.

Where normal filtering -remember, based upon upsampling- just upsamples in order to do its work, Arc Prediction "just" creates higher resolution as had the A/D sample rate been higher. This is important, because the start off point is different. Thus, when the start off is higher resolution, the filtering to remove HD from the too rough stepping is not needed anymore. The stepping is not rough to begin with (4 times more fine and sufficiently enough to measure as good as normal filtering).

With normal filtering the starting point is the reconstruction of individual frequencies (as said before) which does certainly not imply a higher resolution (to the sense of reality). And to keep in mind : *if* that were so, it can't be measured because live music data has no reference (where an individual frequency has).

 

Right;

The above certainly won't make things more clear, but it may shed a light on what is actually going on within DACs against what should be going on. So :

 

- What is going on is that a scientific method which is capable of reconstructing individual frequencies (prooven by measurement), does its work by means of the input of those individual frequencies themselves. Simply put think of this like : when the frequency is put on long enough, the system has learned how it really is and thus should be (reconstructed). Or :

Look at the scope of an e.g. 18KHz sine frequency, and while the waveform is completely distorted, the general looks is an 18KHz sine indeed; The distorted part is more vague, and the eye can see what the idea is.

When this is one big mixture of all stacked frequenies like music exhibits, there is no way something can be learned by the system, unless there is a nice clean frequency going on (which btw for small periods of time happens more often than we may think).

 

- What should be going on is that the too rough stepping (and wrongly placed samples in the time domain !) is reconstructed without the help of next or previous wave cycles. Only then there will be no smearing (simply read : no mixture of sample data) and the sound (music !) is as intended.

 

With the chance of understanding what is - and what should be going on a little better now, let me explain about a possible misconception by many :

 

We want upsampling to achieve a higher resolution

 

Just think of video upscaling, and you know what can be achieve here. BUT :

All the upsampling done, no matter in-DAC or offline, is NOT meant to let increase the resolution. It is only meant to let the filtering work, hence remove harmonic distortion from the audio band. "Filtering" is pretentious here, because it is really not filtering at all. However :

Just injecting samples (like with Linear Interpolation) doesn't help a thing, and HD may even be higher because of it. So, there is more going on, and the normally used (in-DAC) software causing the samples to be injected at the proper place - really can be seen as a filter. But keep in mind : all it does is injecting samples (like 3 on 1 at 4x upsampling).

 

So, upsampling is a means the filter needs to let it work. It does not have the objective of higher resolution.

Also : Who perceived real higher resolution from upsampling ? I never did ...

 

But but but

 

With the starting point being creating higher resolution, well ... we indeed receive higher resolution ! haha

So, where Arc Prediction turned everything upside down (changed the starting point), we started off with higher resolution, and coincidentally don't need a filter because of it.

Hmm ...

 

Now, I know, all of you MAC-boozers don't have the opportunity to listen to Arc Prediction (which is just implemented at the PC side !), but those with the experience of decent video upscaling will know what this brings : "no real HD but most certainly a 100 times better than watching SD material on a 2 meter / 7' screen".

 

The effect of this is crazy, and even goes as far as those who are able to listen to it on the proper hardware (which is 24/192 NOS/Filterless) judge it as better than native 24/192. Including me. How come ? I don't know yet, but something must be going on with the "native" 24/192 etc.

 

It also goes as far as everybody (meaning with 1 exception so far :-) judging Arc Prediction as way better even on heavy OS DACs. This should imply its huge impact, which is -somehow- able to overrule the destroying filtering normally applied. Now :

 

When I_S chimes in, it would be -if possible- very nice to hear some reasoning on how the normally applied (mostly sinx) filtering will not be able to do its normal job, which to me looks like something of "because the filtering has been applied already, it won't perform its destroying work".

Of course, the DAC is now fed with 24/176.4 in the first place, so the "base destroying" can't take place anymore. Still I was astounished to hear from everyone that Arc Prediction on OS DACs gives such an improvement. And this, while I told in advance not to use it on OS DACs because of expected anomalies (those ignorants).

 

In the end it is exactly as has been said in that AA thread : once you have optimized everything there is to optimize in/around a DAC, it is the filtering which becomes very profound. You can also say : it is the last resort to huge improvement, so far the most underestimated. I think.

 

Peter

 

PS: This is all not about the analogue filtering which we may need to protect our amps from high frequencies, which is another story and needed or not needed at all.

 

[cfmsp]

 

As always, thanks Peter for your thorough/detailed response.

 

"All the upsampling done, no matter in-DAC or offline, is NOT meant to let increase the resolution."

 

FULL STOP!

 

"It is only meant to let the filtering work, hence remove harmonic distortion from the audio band. "Filtering" is pretentious here, because it is really not filtering at all".

 

Peter, when you say, "really not filtering at all", can you clarify? What is it, if not filtering?

 

"Also : Who perceived real higher resolution from upsampling ? I never did ..."

 

People report improved sound - of course, most often 'improved' would more rightly be stated as 'different', but for the bias of expectations - from upsampling and perhaps assume it is based on (re)creation of additional samples to fill out the curve (as Steve Nugent has stated). He states regularly that you can fill in the missing data.

 

thoughts?

 

 

"once you have optimized everything there is to optimize in/around a DAC, it is the filtering which becomes very profound. You can also say : it is the last resort to huge improvement, so far the most underestimated. I think."

 

and presumably the least well understood?

 

Cool, something to obsess over other than interfaces and clocks! ;0

 

 

thanks again,

clay

 

 

[barrows]

very interesting, thanks for that. I too believe that digital filtering (and up or over, if one prefers, sampling) are not entirely given their due by those discussing various approaches to improving sound here, and on other forums.

I have one question because I am not entirely sure I understand the idea that up/oversampling filters do not improve resolution. I am clear that the filters do not create new, accurate, data points; but if one takes a filter that is able to significantly reduce pre and/or post ringing, and compares it to a filter which has a lot of pre/post ringing, my suspicion (and listening seems to support) is that the reduction in ringing does unmask more detail (details that were presumably obscured by the ringing), and hence achieves an increase in resolution.

 

[PeterSt]

Hi Barrows,

 

I have one question because I am not entirely sure I understand the idea that up/oversampling filters do not improve resolution. I am clear that the filters do not create new, accurate, data points; but if one takes a filter that is able to significantly reduce pre and/or post ringing, and compares it to a filter which has a lot of pre/post ringing, my suspicion (and listening seems to support) is that the reduction in ringing does unmask more detail (details that were presumably obscured by the ringing), and hence achieves an increase in resolution.

 

You make it very difficult for me to say anything with more sense than you just said yourself. This includes the unmasking by the removal of the ringing, which was my very first impression and how I expressed it on my own forum.

 

Let's say that removing something which masks, also is wrong by itself. It may (!) color ...

 

But let me add this one, also part of the first impressions :

 

What I clearly notice(d) is the "dryness" of it all. The shortness. Not a shortness which removes necessary spatiousness, but a shortness which allows for perceiving the real merits of an instruments. I always have bongos and rototoms in mind. You know, instruments with a most short attack. The short attack jumps forward now, and if anything - this will be because of less smear. Pre, post or mixed.

I also said that at last I could recognize a snare drum. This, while it is one of the most recognizeable instruments of them all. However, not if you have a couple in pocession and from the loudspeakers they keep on sounding strange - no matter you can recognize the instrument by itself.

Brushes are another example. And in general, everything which *needs* no smear or otherwise it becomes messy.

 

The smear from ringing is something anyone can think of. How to achieve no ringing while the "filtering" as such is active is another matter. And therefore I say it again : it is nothing else than preventing the common filtering to (need to) jump in which for NOS listeners was half way the solution (no filtering = no ringing). Half, because they were left with heavy HD, and that by itself causes smear just the same (just look at an FFT from NOS/Filterless and you will get that for sure).

 

But as said, I can't make more sense out of it than you already did yourself.

Regards,

Peter

 

[PeterSt]

Allow me for once to thank you for your ever so kind words. Ok, done :-)

 

Peter, when you say, "really not filtering at all", can you clarify? What is it, if not filtering?

 

I'm afraid you caught me a bit on my own ever feelings/thinking about this;

What I tried to express is that "filtering" as such is an active process to the sense of "filtering out". In this case nothing of the kind is in order. It is just a mathematical process which gives a better result for harmonic distortion opposed to without it. You'd even might call it "multiplying" ...

 

People report improved sound - of course, most often 'improved' would more rightly be stated as 'different', but for the bias of expectations - from upsampling and perhaps assume it is based on (re)creation of additional samples to fill out the curve (as Steve Nugent has stated). He states regularly that you can fill in the missing data.

 

This would all be true;

Sound may improve with the common filtering, but it is always a "so so" thing. To me it never improved, and in the end all of my users (ok, those who expressed themselves on this subject) came to the same conclusion.

However, when indeed the missing samples are injected at the right place (and right = right) there is no reason not to speak of genuine increased resolution. *That* improves clearly (but mind the "when indeed").

Of course I should leave it to your own judgement whether this is increased resolution is there or not (with Arc P.) and that's why I suggested this to be difficult for you MAC users because you won't (be able to) listen to it.

Anyway, knowing what hires brings (yea, higher resolution :-) you can just recognize this with Arc P. Before, I didn't experience any other means of upsampling which did that, which to me clearly is about not injecting the samples at the right place at all. As said, it can reconstruct a nice analoguely looking frequency (which also is about samples injected at the right place) but it cannot do that with complex music.

 

I can expect you want me to say more about this, but I'm afraid I can't. Regarding this, please notice that my statements about common filtering being flawed are merely derived from a now existing better means than I could express it in advance without that means. However :

 

You may recall my ever nagging about measuring flawing; well, it now turns out that *what* we measure is flawed (which I sure couldn't guess one year ago). You mentioned Steve ... he hunts (or hunted) for the same ...

 

Thank you,

Peter

 

[cfmsp]

 

 

Peter,

 

I think I'm with you...however, you took me in a different direction than I expected.

 

Perhaps my comments didn't show it, but ... I am/was biased to believe that filling in the points on the curve would NOT result in 'better' resolution, in the sense that it cannot 'create' details that were not recorded. In essence I was disagreeing with Steve.

 

You seem to be saying that you'd before never experienced (or had never believed it would result in improvement, I'm not sure) an increase in resolution either (from filling in the curve) but, now you have with your own Arc Predictor.

 

Do I understand that, right? I just want to clarify, becuase, as I said, it's not the answer I expected.

 

clay

 

 

 

[PeterSt]

Yes, all correct Clay.

Indeed there may have been a small misunderstanding, because I always see Steve raving about (offline) upsampling, which doesn't cut it to my ears. So, my point about him was solely about the measuring and the disbelief it could be right.

 

So indeed, I too never believed filling in the missing points could lead to somewhere, but this was derived from all the software doing it, leading to no real results. Doing it myself, and knowing how the results should be just by means of looking to graphs of it - brought the good results.

 

So indeed, I too never believed filling in the missing points could lead to somewhere

 

... which actually is not true, because I know (wrong, we know it all) that video upscaling can do it too, and really no one will disagree on the far better results than without, no anomalies whatsoever. But this all works by means of very different algorithms and the purpose is ... better resolution -> upsacling (not removal of distortion ... strange eh ?).

 

Sorry for the confusement, which is always on my side (I learned that by now, haha).

Peter