Reconstruction Filters made audible?

[Ghoostknight]

Hello,

i watched this youtube video:

his test setup was basicly sending something out of the DAC back into a ADC and repeating it to see how audible it can get and there was a significant audible difference between 44,1khz and 96khz, so im wondering if this process just made reconstruction filters obviously audible ? the difference heared pretty much lines up with what im hearing if im not upsampling with cd material, some kind of sharpness / harshness going on

imo its worth a listen if you are curious how reconstruction filters sound "amplified"

 

[gregorio]

his test setup was basicly sending something out of the DAC back into a ADC and repeating it to see how audible it can get and there was a significant audible difference between 44,1khz and 96khz, so im wondering if this process just made reconstruction filters obviously audible ?

It’s difficult to tell, he doesn’t go through the exact procedure he used. Without careful analysis (a null test would have been a good idea), it could be any number of things. For example, the difference could just be some fractional difference in timing/alignment causing a cumulative phase effect. At a guess though, it sounds to me more like a cumulative aliasing issue. If that is the case, it would therefore either be due to the anti-image filter in the DAC or the anti-alias (decimation) filter in the ADC. His experiment with the Focusrite would indicate it’s the decimation filter in the ADC, if indeed it is due to filtering and not some other cause.

the difference heared pretty much lines up with what im hearing if im not upsampling with cd material, some kind of sharpness / harshness going on

So you think you’re hearing a difference in an ADC decimation filter looped back 400 times when playing your CDs, despite the fact you’re not even using an ADC once in your chain, let alone 400 times?

imo its worth a listen if you are curious how reconstruction filters sound "amplified"

Huh? Firstly, how do you get to “reconstruction filters” when the video you yourself posted tends to indicate the ADC is the cause and an ADC does not have any reconstruction filters?! Why not say; “it’s worth a listen if you are curious about oil intercoolers”, because ADCs don’t have any oil intercoolers either!

Secondly, the video you posted is NOT a demonstration of an amplified reconstruction filter! It’s of the DAC output looped back into an ADC, then that resultant digital data output through the DAC again and looped back into the ADC again, done 400 times. In other words, on say the 3rd loop-back the ADC is converting the original signal plus the artefacts of 2 ADC and 3 DAC processes. Feeding the artefacts of the same ADC and DAC chain back into that ADC and DAC chain 400 times is completely different to a “reconstruction filter “amplified””!

The video you posted effectively demonstrates pretty much the exact opposite of what you’re suggesting/claiming. Namely, that even 1200 successive up/down sampling processes together with 1600 successive brickwall filter applications only makes a relatively small audible difference and therefore, just one up/over sampling process and two brickwall filter applications (that occurs in a single pass through a typical DAC) would be so far below audibility as to be laughable!

Why post a video which very effectively demonstrates that your apparent claim/suggestion is completely false? Or maybe you’re suggesting that your DAC is roughly 400x worse than an 18 channel in, 20 channel out ADC/DAC which includes 8 good quality mic pre-amps and costs less than $700?

G

 

[Ghoostknight]

some thoughts:

- while the focusrite clarette test definitely shows that the ADC portion is "better" for this type of work you can still clearly hear a degredation imo
- imo we cant exclude the DAC as long there wasnt a comparison between different DACs, so my guess is both ADC and DAC play a role here together, tho maybe the first ADC was a particular bad example, as far i know most preamps also include a subsonic filter which could play another role here
- i hope we can atleast agree that the differences at 400x cycles is "pretty audible" if you have a trained ear, so is my assumption right that 200x cycles would still give you 50% of the audible change? (and therefore 40 cycles 10% ?)
kind of wished he would have included 50/100/150/200 cycles just to compare them going from the original
- im not exactly sure how preinging adds up from filters but i guess every cumulative cycle would add +3db where the inital cycles adds in comparison "more" than that, so the difference from original to the first cycle is probably a bigger difference than the second cumulative cycle to the first

- goldensound made some files where you can compare a single reconstruction filter if you are curious, imo also a single reconstruction can be audible and it has nothing todo with using a "worse DAC" than others...

 

[bigshot]

An ADC and a DAC are completely separate. I’m sure there are bad capture devices. I’ve never found a bad DAC, except for obsolete NOS DACs.

 

[gregorio]

while the focusrite clarette test definitely shows that the ADC portion is "better" for this type of work you can still clearly hear a degredation imo

You’re joking? What “type of work” do you think requires looping back a recording from a DAC to an ADC 400 times? It can happen in professional mixing and mastering that this type of loop-back occurs once or twice, in edge cases possibly as many a 5 or 6 times but again, this test was 400 times! And as a consumer, why would you be doing even one loop-back, let alone 400?

imo we cant exclude the DAC as long there wasnt a comparison between different DACs, so my guess is both ADC and DAC play a role here together, tho maybe the first ADC was a particular bad example, as far i know most preamps also include a subsonic filter which could play another role here

Huh, the difference was drastically reduced using a different ADC, how does that not demonstrate the difference was mainly due to the ADC? And even using your own logic, your “guess”/conclusion is nonsense: If (according to you) “we can’t exclude the DAC as long as there wasn’t a comparison between different DACs”, then you also cannot “guess”/conclude they “both play a roll together” for exactly the same reason, there was no comparison!

i hope we can atleast agree that the differences at 400x cycles is "pretty audible" if you have a trained ear, so is my assumption right that 200x cycles would still give you 50% of the audible change? (and therefore 40 cycles 10% ?)

We can agree that the difference at 400 loop-backs was “pretty audible”, however, it’s still easily recognisable and the difference is astonishingly tiny considering that’s 400 loop-backs, 800 conversion processes! I would have thought the audible difference would have been far greater. Lastly, where does your “40 cycles” (80 conversions) come from? When a consumer is playing a digital audio file why would there be more than one single conversion (from digital to analogue) and, what percentage would 1/800 of a relatively small difference be?

im not exactly sure how preinging adds up from filters but i guess every cumulative cycle would add +3db where the inital cycles adds in comparison "more" than that, so the difference from original to the first cycle is probably a bigger difference than the second cumulative cycle to the first

So if every cumulative cycle “would add +3dB” and there are 400 cycles, how many dB of pre-ringing would that be in total?

goldensound made some files where you can compare a single reconstruction filter if you are curious, imo also a single reconstruction can be audible and it has nothing todo with using a "worse DAC" than others...

According to his own explanation/description of his test, the only difference within the audible band (<20kHz) between the reconstruction filters he tested was dither that was below -100dB. He doesn’t say exactly what the level was, but even if we take a level higher than he is asserting (-100dB), then at a reasonable peak playback level (say 85dBSPL), that dither noise would be at -15dBSPL and there’s only one anechoic chamber in the world where noise at -15dBSPL would be above the noise floor of the room, but of course he wasn’t in Microsoft’s multi million dollar anechoic chamber! And even if he were, that would still be below the noise floor of his amp + transducers, and even if we ignore that problem as well (!) and assume he has some magical amp and transducers, with an effectively non-existent noise floor, it would still be many times below the human hearing threshold anyway!

Instead of using your brain to come up with nonsense, why don’t you use it to learn and understand some simple, basic facts?

G

 

[GoldenSound]

According to his own explanation/description of his test, the only difference within the audible band (<20kHz) between the reconstruction filters he tested was dither that was below -100dB. He doesn’t say exactly what the level was, but even if we take a level higher than he is asserting (-100dB), then at a reasonable peak playback level (say 85dBSPL), that dither noise would be at -15dBSPL and there’s only one anechoic chamber in the world where noise at -15dBSPL would be above the noise floor of the room, but of course he wasn’t in Microsoft’s multi million dollar anechoic chamber! And even if he were, that would still be below the noise floor of his amp + transducers, and even if we ignore that problem as well (!) and assume he has some magical amp and transducers, with an effectively non-existent noise floor, it would still be many times below the human hearing threshold anyway!

Instead of using your brain to come up with nonsense, why don’t you use it to learn and understand some simple, basic facts?

This assumes that the human limit of hearing is exactly 20khz. Which even if we go based off that, many DAC reconstruction filters do roll off treble to some degree below that.

In the video (I'd recommend watching it before criticizing, I went to great effort to control the methodology and provide means for independent and remote validation) I showed that I was able to audibly discern between a quite typical ESS/AKM linear phase fast rolloff style filter and a very high performance one with millions of filter coefficients even though the content under 20khz showed no major difference. BUT, the caveat was that I also found out that I was able to still hear a bit above 20khz (I'm 26) and so being able to hear 21khz in and of itself could be the sole reason, but in order to find out I'd need to redo the test in a couple years once my hearing range had dropped below 20khz.

I didn't say that the low-level differences from a noise shaper were the reason, as with a meaningful magnitude difference in a range which I can currently hear, that would make most sense as to why I could pass the test until shown otherwise.

If just wanting to test whether differences in signal reconstruction at extremely low levels were beneficial then you'd need a different methodology, using the same frequency domain filter to prevent magnitude differences around nyquist, but using different noise shapers to test the limits of how crazy levels of phase accuracy or quantization error reduction may or may not affect things.

Some like Chord Electronics would say that it does, hence the DAVE being designed with a 1-bit noise shaper that can reconstruct signals at -300dB quite happily, but others would say this makes no sense. But without any controlled testing to verify we don't have a conclusive answer.

 

[bigshot]

This assumes that the human limit of hearing is exactly 20khz. Which even if we go based off that, many DAC reconstruction filters do roll off treble to some degree below that.

Oversampling DACs don’t roll off below 20kHz, and even exceptional schoolchildren don’t hear above 22kHz which isn’t even a single note on the musical scale. Besides all that, unless you listen to gamelan gongs, there isn’t much musical content above 20kHz, and that which is there is at a low volume level and is masked by sound in the audible range.

Super audible frequencies are as useful as teats on a bull hog.

P.S. I don’t believe you can hear above 20kHz at a normal listening level. At higher levels your ears might be distorting audibly. Be careful you don’t go deaf trying to prove your super hearing!

 

[GoldenSound]

Oversampling DACs don’t roll off below 20kHz

They absolutely can, it's entirely dependent on the design of the filter. And there examples of both affordable and very expensive DACs rolling off early.

Take the topping D90 III for example. Several of the filters roll off before 20khz

In fact the only filter on the DAC that does actually attenuate fully by the Nyquist freq (slow minimum shown above in lime green) does have a few dB attenuation by 20khz.
Similar things can be seen on more expensive DACs too. Take the Meitner MA3 for example which is also down a few dB by 20khz:

there isn’t much musical content above 20kHz, and that which is there is at a low volume level and is masked by sound in the audible range.

Super audible frequencies are as useful as teats on a bull hog.

The arguments of the content up there being less than useful are certainly valid, especially since most producers/artists are not going to be able to hear it let alone be doing any intentional mixing/mastering in that region specifically. But that doesn't mean differences there are inaudible.

It just means that the answer as to which filter approach is "best" subjectively is debatable and may not necessarily align with which is most accurate from a purely objective standpoint, since one could easily argue that actually reproducing the content in the ~21khz region 'accurately' may just mean you're getting more HF content that the artist/engineer didn't intend to be there or perhaps wasn't even able to hear was there. Others would say that with certain types of music like more direct/less processed recordings, as close to an ideal nyquist reconstruction would be best.

P.S. I don’t believe you can hear above 20kHz at a normal listening level. At higher levels your ears might be distorting audibly. Be careful you don’t go deaf trying to prove your super hearing!

As said, I'm not certain that the magnitude difference above 20khz WAS the reason I'm able to discern between the different filters in the ABX test, simply that it was the most sensible conclusion given the measured differences observed.

The test was done at normal level with real music, so exactly WHY there is an audible difference I'm not 100% certain, but until I redo this test once my hearing range has dropped in the next year or two, the FR difference itself seems the most likely answer. (Unless you'd prefer I say that the quantization noise differences below -150dB are more likely the cause )

 

[bigshot]

if you use the wrong filter, it might roll off. That is user error. DACs without selectable filters shouldn’t roll off below 20kHz. The $8 Apple dongle gives 20 to 20. If you use a filter that rolls off, that’s user error. You shouldn’t need more than one filter with a modern DAC. Just buy a regular inexpensive DAC and use it. It’s all you need.

 

[GoldenSound]

if you use the wrong filter, it might roll off. That is user error. DACs without selectable filters shouldn’t roll off below 20kHz. The $8 Apple dongle gives 20 to 20. If you use a filter that rolls off, that’s user error. You shouldn’t need more than one filter with a modern DAC. Just buy a regular inexpensive DAC and use it. It’s all you need.

Sure, but then with the limited compute power available in a typical DAC chip, you can't keep things flat under 20khz and actually achieve a correct Nyquist reconstruction, hence why all the other filters shown above on the D90 for instance fail to fully attenuate by the nyquist frequency and will therefore have aliasing.

This is why some products may implement additional compute power in order to be able to use higher performance reconstruction filters that can achieve a full attenuation before Nyquist without affecting content under 20khz.

 

[bigshot]

The $8 Apple dongle is stone flat under 20kHz. I have no idea what you’re talking about.

 

[GoldenSound]

The $8 Apple dongle is stone flat under 20kHz. I have no idea what you’re talking about.

It's not about under 20khz. If you do not attenuate fully by the Nyquist frequency you will have aliased products remaining. Hence why dacs don't just all use crazy slow filters which would be much simpler to implement in the first place.

Nyquist theorem states you must perfectly band limit in order to accurately reconstruct. In practice this isn't possible since a true instantaneous and infinite attenuation would require infinite compute power, so we just do what we can instead and pick the tradeoffs between in-band FR, aliasing rejection and compute requirements that we or a given manufacturer feels are appropriate (or offer a few filter options so the user can choose for themselves)

 

[bigshot]

That stuff would be way below the dynamic range of the music. You’re worrying about inaudible stuff.

 

[GoldenSound]

That stuff would be way below the dynamic range of the music. You’re worrying about inaudible stuff.

And you're assuming it to be inaudible despite existing evidence contradicting that.

(See my video for instance, or have a read through this: https://www.aes.org/tmpFiles/elib/20241226/18296.pdf )

All18 published experiments for which sufficient data could be obtained were included, providing a meta-analysis involving over 400 participants in over 12,500 trials. Results showed a small but statistically significant ability of test subjects to discriminate high resolution content, and this effect increased dramatically when test subjects received extensive training.

If it truly did not matter, then again, DACs (and ADCs) would all just use an extremely basic and slow filter that is flat under 20khz and be done with it.

 

[danadam]

hence why all the other filters shown above on the D90 for instance fail to fully attenuate by the nyquist frequency and will therefore have aliasing.

Imaging, not aliasing.

and pick the tradeoffs between in-band FR, aliasing rejection and compute requirements that we or a given manufacturer feels are appropriate

My impression is that the most common choice is flat to 20 kHz and full image rejection at 24 kHz. Seems reasonable to me.

Of course some manufacturers may choose roll-off before 20 kHz and full image rejection at 22 kHz but that, to me, would be just a broken product.