I tend to prefer slow roll offs and even the NOS mode - what does this say about my setup and preference? (Conclusion: back to fast linear)

[voodooless]

The questions are (1) what exactly distinguishes the HF noise arising from those spurious frequencies, from the HF information encoded by the samples

Well, NOS by definition is just wrong. The sampling theorem says that the samples value is only valid for that specific point in time. For that reason the correct visual representation of a sample is a vertical lines
with dot on it. NOS makes that value average over the whole sample period. That is just invalid! This coupled by staircases that also gives you lots of harmonics give you hell.

A signally better (not less identical though) way would be to first do correct oversampling, let’s say 8x, then average groups of 8 samples again to get back to the original sample rate, and then feed that though a NOS DAC. My suspicion is that the result would be slightly better.

(2) how does oversampling shift that distinguishing line?

By properly filtering the data. The short version:add X new samples in between the existing ones with value 0, take a FIR or IIR low-pass and apply it to the stream: done! The low-pass needs to have enough stopband attenuation to get rid of the HF crap.

 

[solderdude]

I know that, but it doesn't answer my question. The answer is related to the fact that filtering is an essential part of sampling theory. My question is exactly "how" this comes into play with proper oversampling methods. I need to study this further to articulate that.

By interpolation.

 

[MRC01]

Well, NOS by definition is just wrong. The sampling theorem says that the samples value is only valid for that specific point in time. For that reason the correct visual representation of a sample is a vertical lines ...

Preaching to the choir, brother!

 

[voodooless]

Preaching to the choir, brother!

Did you forget the song text then ?

 

[anphex]

Wait, I thought NOS or "super slow roll offs" are just bypassing the filter section of DAC chips completely through a different routing or even pin out BEFORE the reconstruction filter section.

 

[tuga]

Wait, I thought NOS or "super slow roll offs" are just bypassing the filter section of DAC chips completely through a different routing or even pin out BEFORE the reconstruction filter section.

I think that "Real" NOS is digital-filterless and "Pretend" (AKM) NOS is (according to the RME manual) the filter with the smallest steepness (and therefore affecting treble more than the others, but offers the best impulse response).

 

[xaviescacs]

 

[solderdude]

Wait, I thought NOS or "super slow roll offs" are just bypassing the filter section of DAC chips completely through a different routing or even pin out BEFORE the reconstruction filter section.

Omitting filters is not possible for DS DACs so what they do is simply create a 'filter' that does what it is expected to do (mimic sample hold) so that the resulting waveform is close to what comes out of filterless DACs (OS or NOS).

Needless to say this filter option is only there because customers seem to desire such nonsensical modes. The same is true for all other filter options.
Fortunately, most DAC device and chip manfacturers have an option to at least select a proper steep filter (linear and or min. phase).

Such steep risetimes as test signals can make simply can not and do not exist in the original recording. Creating them during the playback (wave forms that originally did not exist) is simply ... well... you know... not the smartest thing to do. All it does is make an artifical test signal nicer looking. A test signal that cannot ever be recorded in reality and only exists in a digitally generated test file.

Its the users choice to use such silly options.

 

[anphex]

I wish I could show someone. I urge people who have the possibility to compare.
Using EAPO, heavy DAC oversampling and then comparing between fast, slow and NOS filters.

Using 192Khz/24 bit, the maximum my Steinberg UR22 provides I played around with RMAA and expected to see higher THD or anything. Only (expected) thing that came up was the slight 0,5dB drop at 20Khz. The UR22 is probably the limiting factor. Still I expected to see some other difference than the FR. Maybe I should get an ADI-2 or MOTU.

I am not even insisting on the impulse response, it's more like a weird artificial brightness was removed without becoming unclear.

 

[voodooless]

Easy test: you already have EAPO: take the fast filter and add the HF rolloff manually. See if it sounds the same..

 

[anphex]

I tried, but no it doesn't really. When I try to replicate the FR of the NOS mode the effect is far less noticable with fast filter + EQ than when using NOS. Even with a more aggresive artificial EQ drop of -3dB. It gets weirder when you remember that I oversample at 384 Khz where the expected drop at 20Khz should be only half of of 192 khz - probably around 0,25dB.

Edit: @voodooless have you heard a NOS filter with heavy oversampling once or are you basing everything on measurement data/theory? Just asking out of curiosity, it isn't meant to sound baiting.

 

[voodooless]

Edit: @voodooless have you heard a NOS filter with heavy oversampling once or are you basing everything on measurement data/theory? Just asking out of curiosity, it isn't meant to sound baiting.

I have not, what does it matter? It’s about what you hear. What I can or cannot hear is irrelevant in that matter.

Your RMAA test is basically worthless if you didn’t use the same settings that you use to listen with. If it was done without upsampling, the differences with upsampling should be even smaller. I would actually not expect any HF droop in that case.

Really a double blind test is needed to verify this.

 

[solderdude]

Is this about personal preference or performance ?
when upsampling what SRC is used ?
What DAC with which settings is used ?
What exactly is compared ?

 

[voodooless]

@anphex, you said that EAPO would solve the Windows upsampling issues? How so? What’s your source of that info? As far as I can see EAPO just uses the kernel sample rate, so the upsampling already happened before it enters EAPO?

 

[anphex]

@anphex, you said that EAPO would solve the Windows upsampling issues? How so? What’s your source of that info? As far as I can see EAPO just uses the kernel sample rate, so the upsampling already happened before it enters EAPO?

This stuck in my head for a while and I couldn't find any. This, and the prior post from you about the really bad windows audio upsampling stack from archimagos blog made me do a really simple thing:

Throw NOS + Windows 10 upsampling out of the window and just try classic 44,1Khz@24bit with fast linear filter.

Pretty embarrassing to admit that now, but it's the best setting so far. I always thought, "Why not run your sound cards at maximum sample rate in case you ever turn on HiRes files? Everything else will then interpolate cleanly somehow." Thanks to you I now know - this time for sure - wrong thinking. The final nudge was reading the Nyquist sampling theorem.
It will now mercilessly stay at 44.1Khz. I may sacrifice the one or other song with 96Khz or more, but I could notice a difference anyway only between 16 and 24 bit. So all is well.

Thanks to all again! It's actually pretty scary how such firm but unbased beliefs can ruin plenty of your gears theoretical peformance and music experience. I don't know why it took me so long to grasp and tackle this. Damn, just what did I miss all this time Stupid me!

And for those who haven't followed this thread, in summary: use a fast roll off filter and disable Windows 10 oversampling.

Edit: Oh, an now all retro games work fine too. Many of them didn't have sound with 384 Khz.

 

[anphex]

In the context of the new findings (see above), I have now tried the filters again with the "correct" Windows 10 settings. I went with my SMSL SU-9n to Fast Linear to Apodizing, then Fast Minimum and now I've ended up with Slow Minimum.

Subjectively it's the most pleasant. I think the minimal minor distortions in the HF range can be tolerated in view of the generally excellent performance of the SMSL + Purifi. Phase shift is inaudible in my current opinion, but pre and post ringing are. "Apodizing" (strong post ringing, no pre), for example, has strong piano notes quite throttled in their attack. But with complex music the sound was pleasantly smooth. The muddy attack bothered me too much though.

Archimagos survey even showed that with ABX most people prefer minimum phase filters.

 

[MRC01]

... Archimagos survey even showed that with ABX most people prefer minimum phase filters.

IIRC, the preference that Archimago found depended on source material. The guitar/lute music has plucky transients, and a relatively larger % of listeners preferred minimum phase. The piano music has harmonic complexity and more listeners preferred linear phase.