PGGB upsampler: DeltaWave null analysis

[danadam]

Sorry for the stupid question.

Already asked, produced 25 pages so far :

What is the point of upsampling?

Perhaps I need to learn more about digital sampling itself, but I just can't fathom what upsampling is supposed to do? Like from 44.1 to 192 gets me what exactly in the end?

www.audiosciencereview.com

 

[solderdude]

It is not a stupid question.

Every DAC has a reconstruction filter in it.
Some DAC's have questionable ones or even completely lack one (on purpose).

What this (in fact any half decent) upsampler does is create a better reconstruction filter with less HF garbage that effectively replaces a questionable reconstruction filter with a better one as the crappy one is now applied much higher up so became irrelevant.
It does so by calculating sample values between the original samples. It will not recreate what once was and the frequency response does not become wider.
The audibility of this aspect is another matter though.

 

[pkane]

Dear @pkane ,

many thanks for running these tests. They seem self-explanatory. Nevertheless, can someone please explain to me (I am neither an electrical- nor audio engineer) how upsampling a track that is based on 16-bit data is supposed to improve my listening experience? Does the software find the missing bits that got lost during the 16-bit recording/sampling?

Sorry for the stupid question.

Hi @fordiebianco , it's a matter of belief in audiophile circles that higher sample rates are needed for higher-resolution sound. Some of it is based on incorrect understanding of how PCM audio works, some of it on the marketing that's been used by audiophile manufacturers and press by making and reinforcing intuitive (but incorrect) claims.

Almost a century ago, Shannon, Nyquist, and others have demonstrated mathematically that a perfect reconstruction of a waveform is possible given that the PCM sampling rate is twice the largest frequency that you want to capture. What that means is that no matter how many points you insert in between the PCM samples, you'll never get a more "perfect" reproduction, so upsampling for the sake of getting "better sound quality" is mostly snake oil. There is one primary engineering use of upsampling/oversampling and that is to simplify and reduce the cost of the reconstruction filter required by Shannon/Nyquist. That's the one legitimate purpose of using a higher sample rate but it's not what audiophiles normally believe or want to hear.

PGGB seems to pray on those same audiophiles by making claims about the apparent need for higher precision in the upsampling conversion. What I demonstrated in the OP is that it really makes no difference at all in any practical sense, since the enhanced precision is way below anything that we could ever possibly hear or any existing DAC could possibly reproduce. Another way to say that it does nothing audible or remotely useful for a hefty price in $ and in effort + complexity.

 

[fordiebianco]

It is not a stupid question.

Every DAC has a reconstruction filter in it.
Some DAC's have questionable ones or even completely lack one (on purpose).

It does so by calculating sample values between the original samples. It will not recreate what once was and the frequency response does not become wider.
The audibility of this aspect is another matter though.

So when I play a redbook CD, what I hear is not the original sample, but the DAC's interpretation of it, via the 'reconstruction filter' ?

 

[pkane]

So when I play a redbook CD, what I hear is not the original sample, but the DAC's interpretation of it, via the 'reconstruction filter' ?

Probably the most shared video on ASR that should help explain how PCM works:

 

[solderdude]

So when I play a redbook CD, what I hear is not the original sample, but the DAC's interpretation of it, via the 'reconstruction filter' ?

Not exactly. Most DACs are multibit Delta Sigma so upsample the incoming data themselves multiple times anyway.

One can never hear the actual samples as these were single points in time at a 44.1kHz interval.
A DAC calculates the values between the samples and produces them. So it is not an interpretation but calculated values.
After that occurred there is a 'post filter' that ensures there is a smooth analog waveform that resembles the original (but band limited) waveform before it was sampled.

Its pure math not an interpretation or estimation as in 'hmm I think this value might be about the correct values'

 

[fordiebianco]

Looks like I have some reading to do. Thank you for your patience and indulging an electronics novice.

 

[PeteL]

Hi @fordiebianco , it's a matter of belief in audiophile circles that higher sample rates are needed for higher-resolution sound. Some of it is based on incorrect understanding of how PCM audio works, some of it on the marketing that's been used by audiophile manufacturers and press by making and reinforcing intuitive (but incorrect) claims.

Almost a century ago, Shannon, Nyquist, and others have demonstrated mathematically that a perfect reconstruction of a waveform is possible given that the PCM sampling rate is twice the largest frequency that you want to capture. What that means is that no matter how many points you insert in between the PCM samples, you'll never get a more "perfect" reproduction, so upsampling for the sake of getting "better sound quality" is mostly snake oil. There is one primary engineering use of upsampling/oversampling and that is to simplify and reduce the cost of the reconstruction filter required by Shannon/Nyquist. That's the one legitimate purpose of using a higher sample rate but it's not what audiophiles normally believe or want to hear.

PGGB seems to pray on those same audiophiles by making claims about the apparent need for higher precision in the upsampling conversion. What I demonstrated in the OP is that it really makes no difference at all in any practical sense, since the enhanced precision is way below anything that we could ever possibly hear or any existing DAC could possibly reproduce. Another way to say that it does nothing audible or remotely useful for a hefty price in $ and in effort + complexity.

That may be all true, and I am not telling people to go and purchase this, and I get that this is different, but still, aren't all DACs that measure well in these pages rely on some form of upsampling? Dacs that rely on simple Nyquist based sampling, (sampling at twice the sample rate) all measure poorly, likely for the reason you mention, difficulty to get a good filter and having it in the analog domain. Delta Sigma is based on oversampling, not on Nyquist. To be fair I do not see straight bogus claims on their site, some questionable yes, but they claim transparency, and mathematically exact conversion. For sound quality improvement it is Indeed questionable, well DSD is questionable period, but at least they don't say, this will sound better. They say with these specific DACs, users have mention better sound. And it's a short lists of DACs. I can't verify that of course. To be honest if this was targeted at mastering engineers I would have no gripes at all. Of course when they suggest to convert your whole collection. They are not talking to professionnal and it is still true that many will spend this 1000 bucks for an improvement that is just in their mind.

 

[Tks]

Dear @pkane ,

many thanks for running these tests. They seem self-explanatory. Nevertheless, can someone please explain to me (I am neither an electrical- nor audio engineer) how upsampling a track that is based on 16-bit data is supposed to improve my listening experience? Does the software find the missing bits that got lost during the 16-bit recording/sampling?

Sorry for the stupid question.

Depends on if there's some interpolation going on.

Upsampling to me seems at best what would happen if you took a 30FPS video, and exported it in a 60FPS encoded container. If there's no interpolation, it looks identical to the 30FPS video (as it would simply frame double). If there is a resampling where interpolation occurs, you'll have newly generated frames. Idk what possible difference this can make once we go above 48kHz (maybe someone can chime in here with an auditory study that demonstrates we can discern each singular Hz of audio sampling if each sample is contained within 96,000 samples per second if we're going to perform a 2x up-sample from 48kHz to make things simple, I simply doubt we're capable of this).

Sure when you evaluate the file you'll see the graph exhibiting less stair-stepping between each sample if you zoom in deep enough. But at the end of the day, all this would be doing at best, is generating potentially falsely generated information that was never present within the original recording (or whatever the final output file format was I should say to be pedantic). There can never be this sort of "lost information recuperation" even if it ends up being the case that the audio file itself might look on-paper closer to the original file recording.

It logically cannot even result in anything other than this, otherwise that would mean upsampling beyond the original recording sampling rate is somehow "recovering lost information" which was never recorded to begin with. That just sounds insane, especially once you start talking about digitally synthesized audio samples (with largely digitally composed music) that aren't "recorded" by anything in the first place to begin with.

 

[pkane]

That may be all true, and I am not telling people to go and purchase this, and I get that this is different, but still, aren't all DACs that measure well in these pages rely on some form of upsampling? Dacs that rely on simple Nyquist based sampling, (sampling at twice the sample rate) all measure poorly, likely for the reason you mention, difficulty to get a good filter and having it in the analog domain. Delta Sigma is based on oversampling, not on Nyquist. To be fair I do not see straight bogus claims on their site, some questionable yes, but they claim transparency, and mathematically exact conversion. For sound quality improvement it is Indeed questionable, well DSD is questionable period, but at least they don't say, this will sound better. They say with these specific DACs, users have mention better sound. And it's a short lists of DACs. I can't verify that of course. To be honest if this was targeted at mastering engineers I would have no gripes at all. Of course when they suggest to convert your whole collection. They are not talking to professionnal and it is still true that many will spend this 1000 bucks for an improvement that is just in their mind.

So they are selling a product that doesn't make any difference compared to a simple and free upsampler. A product that costs $1k+ and is a pain to use and requires a major upgrade in computer hardware just to operate? Isn't that just the textbook definition of a 'true audiophile' product: expensive, hard to use, solves a non-existent problem, and in reality makes no difference?

 

[PeteL]

So they are selling a product that doesn't make any difference compared to a simple and free upsampler. A product that costs $1k+ and is a pain to use and requires a major upgrade in computer hardware just to operate? Isn't that just the textbook definition of a 'true audiophile' product: expensive, hard to use, solves a non-existent problem, and in reality makes no difference?

Again. I am not saying it's worth it. Just that there was 2 things in your statement. You first state that It's been proven that twice the sample rate is proven perfect. While this is true, all modern dacs manufacturers use a technology that upsample, those who don't, like R2R Dacs release products that measure poorly.
Now what you demonstrated is that the upsampler in Deltaware is just as good in practical sense. That's great it tells people to not waste money this extra accuracy don't translate in difference that are audible. Compare to a freeware one. But Nyquist theory tells us nothing about that. The DAC that most of use already upsample and measurments tells us that it do so properly, if your DAC measure well. I just want to clarify, as it may be perceived, by bringing PCM and Nyquist to the conversation, that oversampling is a foolish concept that has no benefit. Maybe in theory but not in the real world, Nyquist based DACs are effectively obsolete. Upsampling is in the conversion chain

 

[pkane]

Again. I am not saying it's worth it. Just that there was 2 things in your statement. You first state that It's been proven that twice the sample rate is proven perfect. While this is true, all modern dacs manufacturers use a technology that upsample, those who don't, like R2R Dacs release products that measure poorly.
Now what you demonstrated is that the upsampler in Deltaware is just as good in practical sense. That's great it tells people to not waste money this extra accuracy don't translate in difference that are audible. Compare to a freeware one. But Nyquist theory tells us nothing about that. The DAC that most of use already upsample and measurments tells us that it do so properly, if your DAC measure well. I just want to clarify, as it may be perceived, by bringing PCM and Nyquist to the conversation, that oversampling is a foolish concept that has no benefit. Maybe in theory but not in the real world, Nyquist based DACs are effectively obsolete. Upsampling is in the conversion chain

Oversampling is an engineering solution to a cost/complexity problem. I've explained why it's useful already. I'm not at all sure why you think that Nyquist somehow doesn't apply to delta-sigma -- it's a mathematical theorem that applies to all sampled data at all fixed rates, oversampled or not.

 

[PeteL]

Oversampling is an engineering solution to a cost/complexity problem. I've explained why it's useful already. I'm not at all sure why you think that Nyquist somehow doesn't apply to delta-sigma -- it's a mathematical theorem that applies to all sampled data at all fixed rates, oversampled or not.

The theorem remains true, tin the sense that a 44.1k files contain all the information to recover a 22k band limited signal, right. But what I mean is that if this file was made using Delta sigma ADC and is recovered by a Delta sigma DAC. It was not sampled at twice the frequency you wanted to reproduce as the theorem ask for, that's what I mean that they are not Nyquist based. It's not just the smoother analog filter that it allows, it's a whole different conception of sampling. In other words, If a Delta sigma ADC with a 5 or so bit Delta representation was only sampling at 44.1K. It could never reproduce 22kHz.

 

[Blumlein 88]

The theorem remains true, tin the sense that a 44.1k files contain all the information to recover a 22k band limited signal, right. But what I mean is that if this file was made using Delta sigma ADC and is recovered by a Delta sigma DAC. It was not sampled at twice the frequency you wanted to reproduce as the theorem ask for, that's what I mean that they are not Nyquist based. It's not just the smoother analog filter that it allows, it's a whole different conception of sampling. In other words, If a Delta sigma ADC with a 5 or so bit Delta representation was only sampling at 44.1K. It could never reproduce 22kHz.

Delta Sigma uses a method to get a result more closely aligned with theoretically possible results. The digital filtering and other things going on let a very inexpensive chip provide state of the art performance of a system based upon 44.1 khz sampling and 24 bit depth. One which no R2R non-oversampling DAC can physically manage to accomplish. You are talking about a difference that makes no difference.

 

[PeteL]

Delta Sigma uses a method to get a result more closely aligned with theoretically possible results. The digital filtering and other things going on let a very inexpensive chip provide state of the art performance of a system based upon 44.1 khz sampling and 24 bit depth. One which no R2R non-oversampling DAC can physically manage to accomplish. You are talking about a difference that makes no difference.

I totally agree with all you say, but I don't understand your last sentence "You are talking about a difference that make no difference" What differences are you implying in those two case, the difference that I am talking about, and the difference that you think I presume it makes?

 

[Blumlein 88]

I totally agree with all you say, but I don't understand your last sentence "You are talking about a difference that make no difference" What differences are you implying in those two case, the difference that I am talking about, and the difference that you think I presume it makes?

That you think one is Nyquist based and one is not. Both are. Two different applications, but neither violate any Nyquist parameters.

 

[PeteL]

That you think one is Nyquist based and one is not. Both are. Two different applications, but neither violate any Nyquist parameters.

OK, I do know that Nyquist is respected. To be honest when I like to understand and learn about engineering concepts I often like analog devices application notes, they are usually well written. In this one they clearly put in opposition oversampling system to Nyquist based sampling, in the end it's just semantics, but bottom line over sampling means just that...

https://www.analog.com/media/en/training-seminars/design-handbooks/system-applications-guide/Section14.pdf

 

[Blumlein 88]

OK, I do know that Nyquist is respected. To be honest when I like to understand and learn about engineering concepts I often like analog devices application notes, they are usually well written. In this one they clearly put in opposition oversampling system to Nyquist based sampling, in the end it's just semantics, but bottom line over sampling means just that...

https://www.analog.com/media/en/training-seminars/design-handbooks/system-applications-guide/Section14.pdf

Noise shaping, digital filtering and decimation are used to reduce the effective sampling rate at the output according to your linked applications guide.

 

[pma]

Dear @pkane ,

many thanks for running these tests. They seem self-explanatory. Nevertheless, can someone please explain to me (I am neither an electrical- nor audio engineer) how upsampling a track that is based on 16-bit data is supposed to improve my listening experience? Does the software find the missing bits that got lost during the 16-bit recording/sampling?

Sorry for the stupid question.

Not the missing bits. If there is a poor reconstruction filter in the DAC used, or no reconstruction filter like mentioned by @solderdude in case of NOS DAC, then upsampling would create more smooth output wave with mirror images pushed upwards (simply put “staircase” output of the NOS DAC would be smoothed to tinier steps). I hate what I wrote in parentheses but cannot find the way how to put it simple.
Resolution remains the same after upsampling. No new information created.

 

[pma]

Noise shaping, digital filtering and decimation are used to reduce the effective sampling rate at the output according to your linked applications guide.

You must be joking, Sir, or my English is not good enough to understand what you are saying. The AD AN is excellent.