NOS DAC sound more natural

[Miska]

Oversampling is lossy, turns out it's impossible to modify audio via lossless methods, so any change, oversample/downsample/EQ/crossover are all lossy.

 

Now the question is, how do you find recordings that have not been made using oversampling ADC?

[Miska]

I actually have quite a few hi-res recordings that sound worse than their 44.1khz counter part. We live in the digital dark ages...

 

Point is that there is very limited amount of content made using ADC conversion section running at 44.1k. Making good quality ADC where conversion section would be running at 44.1k is practically impossible because Nyquist frequency is too close to the audio band. For this reason all modern ADCs are oversampled and the ringing is specifically introduced when rate is reduced either inside oversampling ADC or later at mastering stage. Using DAC with half-band oversampling filter doesn't change this ringing or introduce new ringing. Using DAC with apodizing oversampling filter, the ringing introduced by ADC or mastering stage can be changed/removed. Using a NOS DAC naturally doesn't change the ringing, it is played as-is.

 

Using DSD end-to-end allows A/D and D/A conversions without intermediate sampling rate conversions and specifically allows to avoid rate down conversions which introduce ringing. If the target sampling rate is high enough that there is nothing for the down conversion filter to remove, then no ringing is introduced either, DXD (352.8k) is usually such rate.

[Miska]

So if this aliasing happens above the Nyquist frequency of Redbook, wouldn't we not hear it?

 

Intermodulation may make those frequencies audible...

 

Depending on analog reconstruction filter (or lack of) on a NOS DAC, those image frequencies can span far to MHz range.

 

Whole point of oversampling is to help those analog reconstruction filters remove more of the image frequencies.

[Miska]

I'm curious, is there any evidence to support that NOS is a better fit for 24/96 and up "high res" files? That maybe there is so much information in those files, the need for oversampling just doesn't exist? I'm talking out of my depth here, admittedly. And regardless of the answer, I do prefer redbook through this new DAC than through my old D-S so my question is merely a talking point, not really seeking a point to prove.

 

If you download 176.4/192k hires files, they are effectively 4x "oversampled" already. If you download 352.8k DXD files, they are 8x "oversampled" which is equivalent of the oversampling ratio in most oversampling DACs. Difference is mostly that there is actually also content in the higher frequencies too, compared to something that has been converted from lower sampling rates.

 

So more hires you go with, less you need oversampling and more you sort of close gap to the oversampling DACs.

[Miska]

What you failed to realize is that you can't mix DSD and all sound goes through 44/48khz during mixing (96khz if you are lucky, or unlucky if they are 44->96 then back). Then they upsample it again and that is what you have. Those DSD recording sounded horrible as a result versus PCM in practice.

 

Yes you can mix DSD/SDM, if I can do it (I do it for things like 5.0 channel to stereo downmix), others can do it too.

 

No it doesn't. Sonoma workstation processes DSD at native rate. Merging Pyramix allows editing in DSD and mixing and more complex operations at DXD (352.8 kHz rate).

 

Do you realize that practically all new PCM material is recorded using oversampled sigma-delta ADCs and the PCM you get is down-conversion of DSD-like data coming out of the actual A/D conversion stage?

[Miska]

Do you have any sort of idea what portion of the RedBook and higher resolution music we listen to was treated in this way at the ADC end of things?

 

If you have RedBook content that was recorded at 44.1k, then practically all that content has gone through oversampling ADC and brick-wall filter. The rest that hasn't, has been recorded most typically at 96k sampling rate using oversampling ADC and then converted to 44.1k at mastering stage. For checking out those there's good database here SRC Comparisons . And the brickwall is needed because 44.1k sampling rate just doesn't leave space for doing antialiasing in any other way.

 

If you do recording and playback in DSD, the whole ringing stuff is not an issue, because you don't have decimation filters on the path and the Nyquist frequency of delta-sigma converters is so high, that first order analog antialiasing filter at ADC side is enough to avoid aliases. First order filters don't "ring"... And even if the editing has been done in DXD, the Nyquist of that is still high enough at 176.4k that there is very unlikely content in the source that would trigger the ringing.

 

So the whole ringing thing is only an issue for low rate PCM, and in practically all cases it is built into the content at ADC/mastering stage. It can be later changed with use of apodizing upsampling filters at playback time. But that's it. This is very straightforward and logical thing.

 

When such low rate PCM content is played through a NOS PCM DAC, it just plays the ringing as it is in the source content and produce lot of distortion as a side effect. So the point of avoiding ringing by using a NOS DAC is completely moot (it's a myth)...

[Miska]

Analog filter is always needed for reconstruction... Oversampling/upsampling digital filters are there to help the analog filters. By moving the images that the analog reconstruction filter is supposed to remove, apart from the  actual signal. Otherwise there is really not enough space for the analog filter's transition band.

 

If you want accurate reconstruction of 16-bit digital source, you need to have at least 96 dB attenuation by the first image frequency component (for RedBook that is at 22.05 kHz). If you want accurate reconstruction of 24-bit digital source you need to have 144 dB attenuation correspondingly...

 

[Miska]

ZOH (aka SAH aka S/H) has nulls at multiples of sampling rate. If you upsample/oversample using ZOH, the frequency/phase response is still equal to the original rate because you effectively don't do anything, just increase clock rate. ZOH rolls off slowly, so if you run ZOH for 44.1 kHz sampled data, frequency response begins to roll off already at 0 Hz and becomes most notable 10 kHz onwards. This is why a NOS DAC running at 44.1 kHz rolls off early.

 

Typical oversampling DAC chips, when you give them 44.1 kHz input, run digital filters to 352.8 kHz rate, and then go from there onwards using ZOH (because they don't have enough DSP capacity to do better), typically by 16x. For this reason, you can see images of the input signal around multiples of 352.8 kHz in the output.

 

While for example I run digital filters in software all the way up to 22.5792 MHz, because I'm not short of DSP capacity.

 

For example if we look at 0 - 22.05 sweep at 44.1 kHz sampling rate through iFi micro iDSD with filter set to "bit-perfect", the digital filter is disabled and only ZOH is in use:

You can see images around multiples of 44.1 kHz and the roll-offs.

 

If we then set it's internal digital filter to "standard", it runs digital filter up to 352.8 kHz and then ZOH onwards from there:

Now you can see images around multiples of 352.8 kHz. This is already a bit easier for analog filter to remove. But 2nd order analog filter quite doesn't do it yet, so you can see images up to 2 MHz in the output spectrum.

 

For comparison we can look same source, but upsampled to DSD512 rate in software with digital filters doing up to 22.5792 MHz rate:

We can see there are no images left in the output spectrum...

 

[Miska]

9 hours ago, jamesg11 said:

So, if the idsd receives say upsampled 44.1 to 705 from a player - what then occurs re these digital filters?

 

It gives you images around multiples of 705.6 kHz, so 2x better than it can do with it's internal digital filters. Thus also the first image level is also lower because by that frequency, the analog filter has rolled off 12 dB more.

 

 

[Miska]

On 5/24/2018 at 3:36 AM, buonassi said:

If I'm understanding this quote above, Miska, you're saying that there isn't a phase shift at any of the frequencies between 0 and 22.05khz after passing through the S/H stage?  The high frequencies aren't delayed as they would be in a minimum phase filter?  

 

Perhaps this is the case when ZOH is physically performed vs emulating ZOH using a very weak cubic interpolation filter in software.  As I mentioned in a previous post, @Archimago measured the phase response of what he called the "ultralax" filter in software.  He modeled it to have a very similar frequency and impulse responses of what we'd see in a NOS DAC.  The phase was shifted much more so than in a traditional minimum phase interpolation filter.

 

But I do see where this is an emulation, and not necessarily indicative of the actual ZOH process.  Thoughts?  Confirmations? 

 

ZOH is a "linear-phase filter", just like cubic interpolation and linear-phase FIR. Although it is all-pass, so it doesn't filter anything.

 

You cannot "emulate" ZOH with cubic interpolation because it is different process. ZOH is just a sample copy. Even linear interpolation while being a very bad filter, does something while ZOH does absolutely nothing. For example if you have three sample values 2, 4, 6 and want to interpolate it 2x, ZOH yields you 2, 2, 4, 4, 6, 6 while linear interpolation yields you 2, 3, 4, 5, 6, 7. Both are grossly wrong for audio, but you probably see the point...

 

ZOH process doesn't really do anything, it is practically same as not doing anything at all. It is the simplest digital domain phenomenon you can have, because you just copy same sample value as many times as you want. When you do digital domain analysis of that, you just see wider frequency range because sampling rate is higher, but you can calculate the same easily without doing the ZOH process (because it function is well known). When you do analog domain analysis with ZOH it doesn't change anything compared to running at lower sampling rate, because from analog perspective it doesn't change anything.

 

ZOH is same thing as making digital picture larger by copying each pixel value N times in both directions. So it is not interpolation, but just get bigger image where pixels are bigger blocks - so it begins to look like built from Lego blocks.

 

Cubic interpolation is a real interpolation (filter), although not very efficient one, if you do it with image you don't get as blocky output.