Still some spare time before the F1 GP starts.
Here are the three .wav files in original version plus three processed to 192/24 and back to 44.1/16.
There are 6 random numbers used to identify the files.
Dropbox - NOS - Simplify your life
So I'm anxious for getting feedback on these files.
Hans
Here are the three .wav files in original version plus three processed to 192/24 and back to 44.1/16.
There are 6 random numbers used to identify the files.
Dropbox - NOS - Simplify your life
So I'm anxious for getting feedback on these files.
Hans
Marcel, I forwarded your above post to ZB, author of the PGGB resampler, for comment. His response follows below.
Good morning, ZB,
I shared your interesting email messages with my thread. There is a question from one of our more knowledgeable members regarding your last two messages...I suspect that he is distinguishing between aliasing and image bands. With the concept that aliasing manifests within the desired signal band, while images manifest above or below the desire signal band. So, perhaps, this is just an issue of semantics.
As for myself, I presume that you meant how so many oversampling filter implementations (mostly half-band) do not fully suppress the lowest image-band near Nyquist. I believe that his question is about that, and primarily about how pure NOS (with no analog reconstruction filter, only the zeroth-order-hold masking) does not suppress the image-bands even as well as poor OS does. Therefore, if poor OS allows image-band intrusion, then NOS should allow it even worse. I wonder if you might elaborate a bit for us on what you meant by aliasing artifacts in the context of DAC playback?
-Ken
____________________________________________________________________________
Hello Ken,
Yes that is right, it is an issue of semantics. I was a bit loose with it, strictly speaking, I should not have used aliasing and images interchangeably, but yes I was referring to the very first image near nyquist. Let us say I am very prejudiced when it comes to out of band images and those near Nquist are the worst culprits especially when the input is at CD rates.
I agree both simple ZOH and OS DACs with filters retain the images but the differences we are hearing are two parts, one is the difference in reconstruction error between NOS DAC and OS DAC and the other is the extent to which the images are suppressed and the most important of these images (subjectively for me) are the ones right past Nyquist.
What I was implying is poorly implemented OS DACS can do even more damage compared to simplistic NOS implementation but in my opinion OS DACs when implemented well can surpass NOS DACs in both reconstruction accuracy and in suppressing out of band images. And another factor I had not mentioned earlier is the Noise in the audible range. NOS R2R DACs with a very good and well matched ladder can have a low noise floor in the audible range. OS DACs have to oversample to a very high rate so that they can drop the bit depth to 1 or few bits and rely on noise shaping to reduce the quantization noise. Here too the choice and implementation of the noise shapers and the delta sigma conversion process have a bearing on the sound and poor noise shaping can cause a hardening of the treble and reduced depth and resolution.
I will reiterate that I do not say NOS DACs are better but I am not surprised by the audible difference between the two implementations. While NOS DACs take the less is more opporach (and I do think a analog filter after ZOH is better), OS DACs tend to take a more heavy handed approach and some of them end up doing more damage through multiple/cascaded upsampling each with their own poor quality filters, followed by requantization to a lower bit depth using noise shaping that may not be of high quality (single bit converters are the worst culprits in this). The end result of these two approaches lead to the charesteric sound we associate with NOs and Os DACs.
Regards,
-ZB
-------------------------------------
Author PGGB - remastro
Hi Hans, ready to go...... and BTW congrats with Verstappen !! (I still remember his Dad not being so successful)
Attachments
Here I am! Finally I had some time for the up/downsampling test with the files proposed by Hans. I didn't catch his latest post with the converted files, I did that by myself. So, they do sound different! The easiest way to detect is with the upright bass. In NOS mode, starting with the original file it's like the instrument player hits the chords smoothly. At 192kHz, it sounds just a bit sharper like hitting a little harder. Reconverted to 44,1kHz it sounds more aggressive. First impression is that now the bass is more defined but when I tried to follow the whole stage things seemed harsh. Trying this conversion with the track I've posted earlier it makes the trumpet sound annoyingly smeared.
The original file NOS vs OS? OS is very similar if not identical to NOS converted to 192kHz.
I also remembered to test the infamous telarc 1812 overture. The cannons OS vs NOS is loud sound vs blast!
The original file NOS vs OS? OS is very similar if not identical to NOS converted to 192kHz.
I also remembered to test the infamous telarc 1812 overture. The cannons OS vs NOS is loud sound vs blast!
I also used Audacity, but that is not the whole story.
I upsampled to 192/24 and saved the file.
Then I reloaded the 192/24 file to check wether is was indeed 192/24.
The reason I mention this is because Audacity is a bit tricky and may store a file in a different way as was instructed.
After having reloaded the the 192/24 I resampled it to 44.1 and reduced the bitdepth to 16bit including dither.
So to compare various opinions it may be better to use the 6 files that I made.
Hans
I concur, Hans. Your prepared files reduce the effort for other members to conduct your experiment, making it more likely they will participate. I do think it is important, however, for us to include comparison of different FIR filter engine designs (Windowed-SINC, Equiripple, etc.). Therefore, what might be the best approach all the way around is if we extended your idea to include, at least, two differently processed sets of the same test files. Each set converted using different resampling software, meaning, different FIR filter engine designs.
What do you think?
Ken,
Exactly my idea.
And I think that randomizing the files creates an extra level of confidence in the results, preventing prejudices.
So the files should be offered as .wav files
But first of all, I’ll try to find out the specs of the audacity filters by offering an impulse response and converting up and down dithered silence and see what this brings.
Hans
Exactly my idea.
And I think that randomizing the files creates an extra level of confidence in the results, preventing prejudices.
So the files should be offered as .wav files
But first of all, I’ll try to find out the specs of the audacity filters by offering an impulse response and converting up and down dithered silence and see what this brings.
Hans
Looks to me the prize at present goes to the Holo May : Review and Measurements of Holo Audio May --- Probably the best discrete R2R DAC | Audio Science Review (ASR) Forum
But R2R DACs are coming out at a rate, I could be out of date.
Just for information,
I prepared an impulse of 1 cycle at 44.1/16.
In Audacity I brick wall filtered the spectrum from 20Khz with a passband of 1Khz.
Spectrum of this file is the left part of the first image below.
Then I up-sampled this file to 192/24, see spectrum in the middle part of the image.
Finally back to 44.1/16 gave the spectrum at the right.
The second image shows the pulse in time, now vertically magnified in a log scale to get a much better view on the small details.
From top to bottom the original file after brick wall filtering, then the 192Khz upsampled file and finally the downsampled version.
At first sight, Audacity is doing a proper job, but Fir Filter length does not seem to be very long.
So a much longer filter would be nice to include in the test.
Hans
.
I prepared an impulse of 1 cycle at 44.1/16.
In Audacity I brick wall filtered the spectrum from 20Khz with a passband of 1Khz.
Spectrum of this file is the left part of the first image below.
Then I up-sampled this file to 192/24, see spectrum in the middle part of the image.
Finally back to 44.1/16 gave the spectrum at the right.
The second image shows the pulse in time, now vertically magnified in a log scale to get a much better view on the small details.
From top to bottom the original file after brick wall filtering, then the 192Khz upsampled file and finally the downsampled version.
At first sight, Audacity is doing a proper job, but Fir Filter length does not seem to be very long.
So a much longer filter would be nice to include in the test.
Hans
.
Attachments
Thanks,
I won't tell the correct answer now, but you are close with 2 out of 3.
Let's see what others will find.
Hans
I just finished my listening session. Interesting to do it this way
I listened with my TEST DDDAC:
Roon
PI3 with Ropieee
FIFOPIQ2
Andrea Clock 5,6MHz
DDDAC1794 with one Deck and standard power supply
AC OPTs
Tube amps 6n30 -> B300 and Gydinel speakers
For starters the differences were not large, I had to listen back and forward a bit and then let it flow while reading (I know when stuff gets good, I look up like "oh, what was that ?"
But if I describe the difference in tracks I liked best (I am NOT trying to guess what is what - I am not an audio analyzer...) it was mostly more qubic meter air behind the speakers - more flowing music, more low level detail. Just more natural in my ears
but again, not playing against the wall or anything like that - but at the end clear to me at least
I liked best 2-fields... 21-Bach... and 5-round...
So if "best liked" is unprocessed (otherwise my NOS universe will break down
)then consequently the "processed" ones would be 17-1-13
Thanks Hans for this interesting test setup
Doede
In addition to the sinc impulse function, we can see that the filter kernel is linear phase. Nice. I believe that Audacity utilizes the open-source SoX utility for floating-point resampling. While most of it's filter implementation details appear sketchy, I do recall reading somewhere that SoX is based on a Windowed-SINC filter engine. I couldn't quite determine from the response graphs whether or not it is also a half-band (-6dB @ Nyquist) implementation.
The PGGB resampler may prove our most ready alternative for obtaining a significantly longer filter to experiment with than SoX.
Ken,
When zooming further into the 192 file, I didn’t get the impression that it was using a half band filter.
The PGGB filter is definitely a big step forward in upsampling, but needs a dedicated processor, cannot transform a file into a .wave as far as I understood and it cannot downsample either.
So this seems the perfect candidate for comparing NOS to OS.
For the time being it would be nice if more candidates with a NOS Dac are willing to do the 3 file test.
Statisticically nothing conclusive can be said so far, which would be a pitty for such an interesting theme.
Hans
When zooming further into the 192 file, I didn’t get the impression that it was using a half band filter.
The PGGB filter is definitely a big step forward in upsampling, but needs a dedicated processor, cannot transform a file into a .wave as far as I understood and it cannot downsample either.
So this seems the perfect candidate for comparing NOS to OS.
For the time being it would be nice if more candidates with a NOS Dac are willing to do the 3 file test.
Statisticically nothing conclusive can be said so far, which would be a pitty for such an interesting theme.
Hans