• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

PGGB upsampler: DeltaWave null analysis

solderdude

Grand Contributor
Joined
Jul 21, 2018
Messages
15,896
Likes
35,916
Location
The Neitherlands
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.

The vast majority of commercial recordings in studios are made at 96kHz or 192kHz. Some even higher. DS or different ADC is immaterial though.
After the recording is completed it is down-sampled to various formats. A very common one being 44.1/16.
This is always done with a sharp digital filter cutting off everything above 20kHz or 21kHz or so.
Dither can be used to effectively increase dynamic range of 16 bit files.
Then there is a product that is sold.
It is this product that has to be reproduced. When one wants > 20kHz there is often a high-res option 96/24 or 192/24 or higher for people that think they need it.
Then there is MQA as well.

This is the only format where a 44.1/16 container also has >20kHz content after decoding at the cost of bit depth. Note that some recordings still can sound very good even with 10 bits, only having a bit more noise when playing really loud.

Upsampling is only beneficial when the hardware it is played on performs technically much better when higher bitrates are used (signal fidelity).
This can even be the case when a physical DAC designer decides to use a DS DAC chip and only uses one of the poorer (not adhering to the sampling theorem) reconstruction filters. In this case it is done on purpose for people that believe this is the way to go.
Weirdly enough upsampling improves technical performance but at the same time 'removes' exactly that aspect that DAC is actually sold for as one still will be listening to upsampled/sharp filtered 44.1 files with no content above 21kHz and including the dreaded 'pre- and post-ringing'.
 

Blumlein 88

Grand Contributor
Forum Donor
Joined
Feb 23, 2016
Messages
20,529
Likes
37,060
You must be joking, Sir, or my English is not good enough to understand what you are saying. The AD AN is excellent.
Yes the AD AN's are fine. My post was nearly a direct quote. The part "are used to reduce the effective sampling at the output" is word for word.
 

PeteL

Major Contributor
Joined
Jun 1, 2020
Messages
3,303
Likes
3,839
Noise shaping, digital filtering and decimation are used to reduce the effective sampling rate at the output according to your linked applications guide.
But is there a debate on that? Not sure what you are trying to make me say.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,635
Likes
10,219
Location
North-East
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.

I think you're misreading that Analog Devices paper. When it says 'Nyquist' vs 'non-Nyquist' they are talking about two different sampling rates (original vs. oversampled), not two different methods or concepts. The sampling theorem remains in effect in both cases. And the reason for oversampling is stated clearly right at the start of the AD application note, and I repeat:

A major advantage of oversampling an analog signal is the resulting simplification in the analog antialiasing filter requirements. The disadvantage of simple oversampling is that it also increases the ADC output data rate which increases the required size of the buffer memory; and in real-time DSP applications, there is less time between samples to perform calculations.
 

PeteL

Major Contributor
Joined
Jun 1, 2020
Messages
3,303
Likes
3,839
The vast majority of commercial recordings in studios are made at 96kHz or 192kHz. Some even higher. DS or different ADC is immaterial though.
After the recording is completed it is down-sampled to various formats. A very common one being 44.1/16.
This is always done with a sharp digital filter cutting off everything above 20kHz or 21kHz or so.
Dither can be used to effectively increase dynamic range of 16 bit files.
Then there is a product that is sold.
It is this product that has to be reproduced. When one wants > 20kHz there is often a high-res option 96/24 or 192/24 or higher for people that think they need it.
Then there is MQA as well.

This is the only format where a 44.1/16 container also has >20kHz content after decoding at the cost of bit depth. Note that some recordings still can sound very good even with 10 bits, only having a bit more noise when playing really loud.

Upsampling is only beneficial when the hardware it is played on performs technically much better when higher bitrates are used (signal fidelity).
This can even be the case when a physical DAC designer decides to use a DS DAC chip and only uses one of the poorer (not adhering to the sampling theorem) reconstruction filters. In this case it is done on purpose for people that believe this is the way to go.
Weirdly enough upsampling improves technical performance but at the same time 'removes' exactly that aspect that DAC is actually sold for as one still will be listening to upsampled/sharp filtered 44.1 files with no content above 21kHz and including the dreaded 'pre- and post-ringing'.
Yes, sorry I feel there are two different conversation stream happening at the same time. I am partly responsible for not being clear. The part you quoted from me was about oversampling in term of when talking about Delta Sigma DACs and ADC. Not about the general benefits or not of High res or the benefit or not of upsampling to high res typical 44.1 files. But Yes my first foray in audio was trough recording Studios. Still the general concept than avoiding the need of sharp analog filter is present in both concepts. "Sharp digital filter", altough, highly suggest an over sampling system like Delta Sigma.
 

PeteL

Major Contributor
Joined
Jun 1, 2020
Messages
3,303
Likes
3,839
I think you're misreading that Analog Devices paper. When it says 'Nyquist' vs 'non-Nyquist' they are talking about two different sampling rates (original vs. oversampled), not two different methods or concepts. The sampling theorem remains in effect in both cases. And the reason for oversampling is stated clearly right at the start of the AD application note, and I repeat:
But I never debated that It was about the simplification of anti aliasing filter. I agreed with you on that from the start. Yes, if you are trying to make me say that with Delta sigma you'll have more than 22.1k or reproduction has the limit Nyquist impose, no, I don't think that, I am just saying that Nyquist frequency on a Delta Sigma is not half the sample rate, that's all, and by sample rate, I mean the rate it samples, not the file it produces to store on your medium. Now if you don't see two different methods or concept, maybe you are misreading the paper? They are related, sure, but they are different methods.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,635
Likes
10,219
Location
North-East
I am just saying that Nyquist frequency on a Delta Sigma is not half the sample rate, that's all
Ok, maybe I'm missing something, but that's the definition of oversampling: it's increasing the original sampling rate, one that satisfied the Nyquist criterion. Why is this even a discussion, then?

Now if you don't see two different methods or concept, maybe you are misreading the paper? They are related, sure, but they are different methods.
Who ever said that they are the same method? I've written a PCM to DSD converter (and DSD to PCM), so I know a little about the differences between the two. I still have no idea what you're trying to say and how any of it applies to the OP analysis.
 

PeteL

Major Contributor
Joined
Jun 1, 2020
Messages
3,303
Likes
3,839
Ok, maybe I'm missing something, but that's the definition of oversampling: it's increasing the original sampling rate, one that satisfied the Nyquist criterion. Why is this even a discussion, then?
Yep, I am starting to wonder why is it even a discussion.
Who ever said that they are the same method?
Pkane: -When it says 'Nyquist' vs 'non-Nyquist' they are talking about two different sampling rates (original vs. oversampled), not two different methods or concepts.
Those are your words
Well Taking an amplitude sample on 16 or 24 bits resolution at a twice the bandwith rate and Encoding the delta in a handful of bits at many times that rate, are two concepts, two method...
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,635
Likes
10,219
Location
North-East
Yep, I am starting to wonder why is it even a discussion.

Pkane: -When it says 'Nyquist' vs 'non-Nyquist' they are talking about two different sampling rates (original vs. oversampled), not two different methods or concepts.
Those are your words
Well Taking an amplitude sample on 16 or 24 bits resolution at a twice the bandwith rate and Encoding the delta in a handful of bits at many times that rate, are two concepts, two method...

And completely unrelated to this topic, so what is your point?
 

PeteL

Major Contributor
Joined
Jun 1, 2020
Messages
3,303
Likes
3,839
And completely unrelated to this topic, so what is your point?
You are the one that brought Nyquist in the discussion, not me!
My point was that bringing the core PCM/Nyquist basic theory suggested that there was no benefit of sampling at higher sample rate than twice bandwith. I was just saying that the test you made don't demonstrate that, And I was putting a counter example to that, noting that all modern DACS output samples at a higher rate than that, in a different form. That's all.
 
Last edited:

PeteL

Major Contributor
Joined
Jun 1, 2020
Messages
3,303
Likes
3,839
Sorry, you’re confused and now we’re going in circles. Either post something related to the topic that’s worth discussing or stop trolling this thread.
Edited my post for some clarity...
 

Matias

Master Contributor
Forum Donor
Joined
Jan 1, 2019
Messages
5,034
Likes
10,814
Location
São Paulo, Brazil
Great test, thanks @pkane. I wonder how the results would be with a widely popular upsampler library like SoX, does the Delta Null also show extremely low level differences?
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,635
Likes
10,219
Location
North-East
Great test, thanks @pkane. I wonder how the results would be with a widely popular upsampler library like SoX, does the Delta Null also show extremely low level differences?

Worth a try. I'm not a sox user, so maybe someone can do that test, or let me know the best parameters to try to upsample from redbook to 705.6k/24 or 32 bit and I can try it here.
 

danadam

Addicted to Fun and Learning
Joined
Jan 20, 2017
Messages
959
Likes
1,498
Worth a try. I'm not a sox user, so maybe someone can do that test, or let me know the best parameters to try to upsample from redbook to 705.6k/24 or 32 bit and I can try it here.
Practically?
Code:
sox INPUT OUTPUT rate 705600
Probably doesn't make sense to aim for better noise rejection than 16 bit.

But I guess we don't care about practicality :) Then one of those:
Code:
sox INPUT -b32 OUTPUT rate 705600
sox INPUT -b32 OUTPUT rate -v 705600
sox INPUT -b32 OUTPUT rate -s 705600
depending which tradeoffs one prefers:
full.png

zoom.png
 
Last edited:
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,635
Likes
10,219
Location
North-East
Great test, thanks @pkane. I wonder how the results would be with a widely popular upsampler library like SoX, does the Delta Null also show extremely low level differences?

Here's the comparison of the same upsample from PGGB used in OP to the same track upsampled by SoX to 705.6k/64 bit:

Code:
sox -D -v0.9 input.flac -b64 output.wav rate -v 705600

RMS of the difference is -152dB / -199dBA (DeltaWave was -187dB/-204dBA):

1677604761953.png


Time-domain difference (null) file:
1677604899589.png


Spectrum of the null file:
1677604986016.png


So, DeltaWave upsample result is closer to PGGB than SoX with -v option, but overall, the difference is still completely inaudible and well below 23 bits.
 
Last edited:

Matias

Master Contributor
Forum Donor
Joined
Jan 1, 2019
Messages
5,034
Likes
10,814
Location
São Paulo, Brazil
Thanks! So the difference around -140 dB is below what even the most high end DACs can resolve.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,635
Likes
10,219
Location
North-East
Oh, I didn't know about that. But the internal processing still seems to be 32 bits only.

Since DeltaWave uses 64-bit FP internally, I thought it would be better to have SoX generate the 64-bit samples.

Now that you mention it, the small difference between DeltaWave and SoX upsamplers is likely due to the higher precision of calculations in DW as it uses 64-bit calculations throughout.
 
Top Bottom