I can agree in that there could be somewhat of a reference problem in these evaluations.... I'm sure that somewhere in thread there is a thesis about how we can evaluate the files that have been presented even if we have different kind of reproduction chains - maybe it could be good to reiterate the description at this time? I myself became uncertain and curious.
//
Hi, lcsaszar,
I should clarify, that there is no guessing with a proper SINC-function interpolation-filter. There are other interpolation methods which I believe do a certain amount of guessing. Such as 'spline' and 'polynomial' based interpolators. Which involve curve fitting through the given sample points. I suppose, this process entails some guessing, but I'm not knowledgable about these methods to say for certain. Marcel probably knows. The sampling theorem, however, requires a SINC based interpolation to support perfect reconstruction of the signal.
Last edited:
TNT, I don't know. For the last test we were told that we should evaluate which samples are carrying pleasure or displeasure after long listening, but based on this criteria I had no clue. I even set a playback in repeat mode for two days and none of these tracks delivered a discomfort nor fatigue which I experienced with DS DACs. Then this post brought me on the right track:
Now it is clear for me that introduced echo is destroying NOS characteristic of sound. I wish if were more NOS users.
In fact a difference was small but noticeable, and I picked up the right sequence on the Saint-Saens track. Here is my post #1075 describing sound properties change: What do you think makes NOS sound different?
Now it is clear for me that introduced echo is destroying NOS characteristic of sound. I wish if were more NOS users.
Last edited:
Hi Ken,
It is clear now. But is the digital filter that results in a sinc function in the analog domain the only reconstruction filter alternative? I mean just like there are a zillion analog filters (Butterworth, Bessel, Chebyshev, etc) there must be other digital filter variants. And I read somewhere that there can be an inverse sinc pre-emphasis in the digital domain, in order to compensate for the subsequent sinc droop. I think even the SAA7220 has it built in.
It is clear now. But is the digital filter that results in a sinc function in the analog domain the only reconstruction filter alternative? I mean just like there are a zillion analog filters (Butterworth, Bessel, Chebyshev, etc) there must be other digital filter variants. And I read somewhere that there can be an inverse sinc pre-emphasis in the digital domain, in order to compensate for the subsequent sinc droop. I think even the SAA7220 has it built in.
That is a good question. It seems that as long as the Nyquist frequency is high enough to allow the filter to fully reach it's stop-band, a more shallow transistion slope filter could be employed. Where a brickwall (SINC) filter comes in to it's own is when the desired signal band and the Nyquist frequency are close to each other. As it is with CD.
Most FIR interpolation-filters additionally include digital-EQ of the SINC droop.
Last edited:
Just for clarity, the sinc filter Ken was writing about has a sinc-shaped impulse response and a rectangular (brick wall) frequency response. A zero order hold has the opposite: rectangular impulse response and sinc-shaped frequency response.
I'm pretty sure polynomial-based interpolation can be seen as a linear low-pass filter with an impulse response that depends on the order of the polynomial. For example, first-order (a.k.a. linear) interpolation corresponds to a triangular impulse response and a sinc^2 frequency response, so expressed in dB twice as much treble roll-off as a zero order hold.
I don't know anything about spline interpolation, but what I've read about it on Wikipedia sounds very non-linear, not like a low-pass filter at all.
The issue with sampling is that it produces copies of the signal spectrum around all multiples of the sample rate. When you restrict the bandwidth to less than half the sample rate (a.k.a. the Nyquist frequency) before sampling, those copies don't overlap each other and don't overlap the desired signal. This is anti-alias filtering. When they don't overlap the desired signal, you can get rid of them when you convert the sampled signal back into a continuous time signal. You do that with another low-pass filter known as the reconstruction filter or anti-imaging filter.
Using brick-wall filters just below the Nyquist frequency gives you the largest possible bandwidth while avoiding aliasing and imaging, namely a bandwidth almost up to the Nyquist frequency. If you can live with less bandwidth because the sample rate is larger than the bare minimum, you can use smoother filters. For example, this filter is meant for 20 kHz audio bandwidth at 88.2 kHz sample rate.
Attachments
Last edited:
Given that my IQ is below the top of the bell curve and that your knowledge and incite on this topic seems extraordinary in the magnitude of my light I have some questions Marcel.
Is it reasonable to conclude that under circumstances of absolute transfer linearity of numeric values to a corresponding voltage output of a power amplifier, this along with changing voltage steps in zero time, that no form of filtering is required to achieve perfect reproduction of an NOS signal at the output of that amplifier?
What I am trying to get at is that it seems the degree of filtering is only necessary as a function of the ability of a sequence of following devices to handle the remaining bandwidth at their inputs. Hence a slow rolloff filter, as perhaps a polynomial type, would be adequate if subsequent networks can adequately filter out those images before being imposed to further devices down the land, ending up to the input of a power amplifier. A power amplifier seems often the device most limited in handling an unfiltered NOS input.
This further suggests that attaching a volume control to the output of a polynomial type filtered NOS signal, being then connected directly to the input of a power amplifier, has variant consequences whereupon flame retardation might be in order. Ultimately brick wall filtering seems only useful as a safeguard if the nature of following networks is unknown.
Gerrit
If there is nothing further down the chain that gets disturbed in some way or other by the ultrasonic copies - no amplifiers or loudspeakers that can't handle the ultrasonic hash properly, no cat who gets irritated by them, nothing - then you can leave out the reconstruction filter. You are basically using the bandwidth limitations of your ears as reconstruction filter then.
A typical NOS DAC has a zero-order hold function that already droops to -3.1678 dB at 20 kHz at 44.1 kHz sample rate. If that bothers you, you may need some treble boost to correct for it. With linear interpolation, it would be -6.3356 dB at 20 kHz.
A typical NOS DAC has a zero-order hold function that already droops to -3.1678 dB at 20 kHz at 44.1 kHz sample rate. If that bothers you, you may need some treble boost to correct for it. With linear interpolation, it would be -6.3356 dB at 20 kHz.
Cats certainly get annoyed by THD distortion in this range, but:
Who knows whether cats get annoyed by a presence of coherent images?
A part of Sony propaganda when introducing SACD was promoting a wide bandwith amplifiers and speakers. Many years passed since SACD died and there are still positive aspects of this campain.
Who knows whether cats get annoyed by a presence of coherent images?
A part of Sony propaganda when introducing SACD was promoting a wide bandwith amplifiers and speakers. Many years passed since SACD died and there are still positive aspects of this campain.
Last edited:
This was expected Marcel. To expand, if a filter was introduced on the input of this amplifier of the kind in your thumbnails of post #1287 this seems would satisfy the SPCA and address the other things outlined. The conclusion is that there seems no sonic advantage in performing any digital or analog filtering of an NOS signal unless some following device is incapable of adequately rejecting the amplitudes and bandwidth of images contained on its input. This is to suggest that the sonic merits of various digital filters tested or proposed is a compromise between the negative aspects of including a particular digital filter and the need for it, being a function of the analog system following it to accomplish that task.
There is also a good reason to filter if you care about the substantial treble loss of a zero order hold. If you should want to equalize out the treble loss without causing phase errors, you have to do it with a linear-phase equalizer, while most analogue tone controls are minimum rather than linear phase.
Last edited:
A "Substantial" treble loss, I can't hear. I wouldn't notice on A/B tests, unless filters are degrading a sound natural properties, like (for me) the last test has proven, or would bring a digital 'glare' to a sound.
Lets younger ones judge (those who can hear above 12kHz).
Marcel, did you edit your last statement about 'cats'?
Lets younger ones judge (those who can hear above 12kHz).
Marcel, did you edit your last statement about 'cats'?
Back from holiday, here are the long awaited 176.4/24 versions.
I include the original 44.1/16 and the already auditioned 88.2/24 versions.
Objective is to find out whether you can hear a difference between the 88.2/24 and the 176.4/24 and if so, which of the two you prefer by what margin.
When not sure about the differences you hear, you can go back to 44.1/16 as the reference.
Also for those who can compare will it be interesting to receive your information whether the sound between NOS and OS has become closer when using this 176.4/24 files compared to the lower bitrates.
Some background of what ZB has done for us with great care with his PGGB software:
The original 44.1/16 files were up-sampled and FIR filtered to 176.4 in 64 bit.
The higher bitrate made it possible to perform noise shaping while quantizing and dithering to 24 bit as the next step.
The noise shaping process reduced the newly added noise in the audio range, coming from all calculations and quantizing, to -270dB or in other words, no additional noise was added to the original content resulting in very clean versions.
Please PM me with your findings, while mentioning whether you were using a NOS or OS Dac and the margins of the auditioned differences
Hans
.
Dropbox - NOS-44.1 - Simplify your life
Dropbox - NOS1_88.2 - Simplify your life
Dropbox - NOS_176_24bits - Simplify your life
I include the original 44.1/16 and the already auditioned 88.2/24 versions.
Objective is to find out whether you can hear a difference between the 88.2/24 and the 176.4/24 and if so, which of the two you prefer by what margin.
When not sure about the differences you hear, you can go back to 44.1/16 as the reference.
Also for those who can compare will it be interesting to receive your information whether the sound between NOS and OS has become closer when using this 176.4/24 files compared to the lower bitrates.
Some background of what ZB has done for us with great care with his PGGB software:
The original 44.1/16 files were up-sampled and FIR filtered to 176.4 in 64 bit.
The higher bitrate made it possible to perform noise shaping while quantizing and dithering to 24 bit as the next step.
The noise shaping process reduced the newly added noise in the audio range, coming from all calculations and quantizing, to -270dB or in other words, no additional noise was added to the original content resulting in very clean versions.
Please PM me with your findings, while mentioning whether you were using a NOS or OS Dac and the margins of the auditioned differences
Hans
.
Dropbox - NOS-44.1 - Simplify your life
Dropbox - NOS1_88.2 - Simplify your life
Dropbox - NOS_176_24bits - Simplify your life
176.4kHz Files Released
Don't miss the release of the 176.4kHz PGGB upsampled files via the Dropbox link located at the bottom of Hans' post #1295 above. As Hans indicates, this experiment will test for your preference, or, perhaps, lack of preference between these 176.4 upsampled files, and their PGGB 88.2 upsampled versions and/or the 44.1 source files played NOS. If possible, listen via the same DAC which you utilized for our prior PGGB 88.2 upsampling experiment. Be sure to indicate in your report to Hans whether a given preference was versus 88.2 upsampled file playback, or versus 44.1 NOS playback. In fact, reporting separate preferences versus each of these two previous file groups would be great, should you have the patience to provide them.
Also, indicate whether you listened via an NOS DAC, or an OS DAC. Most importantly, have fun.
Don't miss the release of the 176.4kHz PGGB upsampled files via the Dropbox link located at the bottom of Hans' post #1295 above. As Hans indicates, this experiment will test for your preference, or, perhaps, lack of preference between these 176.4 upsampled files, and their PGGB 88.2 upsampled versions and/or the 44.1 source files played NOS. If possible, listen via the same DAC which you utilized for our prior PGGB 88.2 upsampling experiment. Be sure to indicate in your report to Hans whether a given preference was versus 88.2 upsampled file playback, or versus 44.1 NOS playback. In fact, reporting separate preferences versus each of these two previous file groups would be great, should you have the patience to provide them
. Also, indicate whether you listened via an NOS DAC, or an OS DAC. Most importantly, have fun.
Last edited:
In my mind, this kind of test is flawed in two ways. Please bear with me, I will try to provide an alternative way to conduct the tests.
First off, the flaws:
1) This test cannot or is not intended to be conducted as a blind test. This means two things: People with an agenda could cheat willingly. And secondly I think our brain is very susceptible to expectation bias. People that don't want to cheat might still be influenced by their excpectations.
Don't get me wrong, I don't want to accuse anybody of wanting to cheat on this test.
2) If you use a DAC with no resampling at all, the DAC will run at different sampling rates during the test which in itself might introduce more or less severe alterations to the perceived sound (remember the zero order hold droop which will shift by an octave or two by going from 44 to 88 or 176?)
So here is my proposal:
a) Use an original file which has between 44k and 96k FS (you will see why later).
b) Reduce the level of the original file by ~6dB and dither it properly (to avoid intersample over issues), offer that as option 0 on the test.
c) Upsample the file produced in b) by the resampling algorithms to be tested to 176k or 192k, you might even deliberately want to use a non integer upsampling ratio because using a non integer upsampling ratio is one of the things ANY resampling algorithm HAS TO DO CORRECTLY, otherwise it is flawed IMO.
d) Downsample files produced by c) to the original sample rate, offering files 1 to N as test files.
This way we would produce N + 1 files which use the resampling algorithms twice while producing files of the same resulting sample rate. N being the number of resampling algorithms we want to test.
Nobody would be able to guess which file is which simply by looking at the file size or the FS the DAC is running at. Also expectation bias would be ruled out. And DACs changing their sound signature at different FS would also be ruled out.
There could then be a single person conducting the preparation of the files, gathering of the results and he/she could present the results in a similar manner that Marcel has done for the last test round, the echo listening test. (Which he did an a superb manner I might add, hats off to him and thanks for his dedication and time)
TLDR: I consider non-blind listening tests of this kind to be irrelevant. So if the goal of this thread is to be scientific - well, then let's be scientific.
Tfive,
Please provide a link to a peer-reviewed published study (1) defining 'expectation bias' as you have described it, and (2) that shows the bias has been demonstrated to affect results in the type of listening test being conducted here.
I just ask seeing as how you you indicate you to want to see a scientific test, so please let's first review the scientific basis of your underlying premise. Thanks.
Please provide a link to a peer-reviewed published study (1) defining 'expectation bias' as you have described it, and (2) that shows the bias has been demonstrated to affect results in the type of listening test being conducted here.
I just ask seeing as how you you indicate you to want to see a scientific test, so please let's first review the scientific basis of your underlying premise. Thanks.