Ken, I wanted to add something to the reply to your question in the other thread. I've copied the question and reply here, because it is closer to the original topic here. The new part is part 3 at the end.
I'm not sure if this answers your question, but:
1. A sigma-delta modulator has to work at a clock frequency much greater than twice the highest signal frequency of interest to get a decent amount of noise shaping, so in that sense it is always oversampled.
2. You can look at a digital sigma-delta modulator as a kind of IIR filter with a very coarse requantization in its loop. The signal transfer from input to output, usually called the STF (signal transfer function), can be an all-pole low-pass transfer or there can be additional zeros in it, depending on the design. One can even cover all poles with zeros and have a unity signal transfer function.
In any case, the poles of the signal transfer function are also the poles of the noise transfer function (NTF), a high-pass transfer function that describes the transfer from quantization noise to the output. They determine at what frequency the NTF levels off, and that has to be far above the audio band to get a decent amount of noise suppression at audio frequencies. It's usually of the order of some hundreds of kilohertz for an audio converter.
All in all, the STF will usually either be an all-pole minimum-phase low-pass transfer with a bandwidth of hundreds of kilohertz, or a minimum-phase transfer that peaks and then rolls off at hundreds of kilohertz, or a perfectly flat response. None of them look anything like a brick-wall linear-phase filter at or just above 20 kHz.
3. A sigma-delta will by itself have no pre-ringing or pre-echo, as the STF is usually either minimum phase or minimum phase except for a constant delay.
Whether there can be intersample overshoot issues depends on the design. There certainly won't be on the ones with flat STF. There may be on the ones with an all-pole low-pass STF, but probably less than on a digital interpolation filter. That is, as the cut-off frequency is far above the audio band, any overshoots will be much shorter than on a low-pass filter that cuts off just above the audio band. Besides, you can't design a sigma-delta to be exactly at the edge of clipping like you can a digital low-pass filter.
I would expect the worst overshoot issues on sigma-deltas that have out-of-band peaking due to zeros in the STF that don't cover the poles. Those are often used in ADCs, but I doubt if anyone ever uses them in DACs. At least I can't think of any advantage of doing so.
I'm not sure if this answers your question, but:
1. A sigma-delta modulator has to work at a clock frequency much greater than twice the highest signal frequency of interest to get a decent amount of noise shaping, so in that sense it is always oversampled.
2. You can look at a digital sigma-delta modulator as a kind of IIR filter with a very coarse requantization in its loop. The signal transfer from input to output, usually called the STF (signal transfer function), can be an all-pole low-pass transfer or there can be additional zeros in it, depending on the design. One can even cover all poles with zeros and have a unity signal transfer function.
In any case, the poles of the signal transfer function are also the poles of the noise transfer function (NTF), a high-pass transfer function that describes the transfer from quantization noise to the output. They determine at what frequency the NTF levels off, and that has to be far above the audio band to get a decent amount of noise suppression at audio frequencies. It's usually of the order of some hundreds of kilohertz for an audio converter.
All in all, the STF will usually either be an all-pole minimum-phase low-pass transfer with a bandwidth of hundreds of kilohertz, or a minimum-phase transfer that peaks and then rolls off at hundreds of kilohertz, or a perfectly flat response. None of them look anything like a brick-wall linear-phase filter at or just above 20 kHz.
3. A sigma-delta will by itself have no pre-ringing or pre-echo, as the STF is usually either minimum phase or minimum phase except for a constant delay.
Whether there can be intersample overshoot issues depends on the design. There certainly won't be on the ones with flat STF. There may be on the ones with an all-pole low-pass STF, but probably less than on a digital interpolation filter. That is, as the cut-off frequency is far above the audio band, any overshoots will be much shorter than on a low-pass filter that cuts off just above the audio band. Besides, you can't design a sigma-delta to be exactly at the edge of clipping like you can a digital low-pass filter.
I would expect the worst overshoot issues on sigma-deltas that have out-of-band peaking due to zeros in the STF that don't cover the poles. Those are often used in ADCs, but I doubt if anyone ever uses them in DACs. At least I can't think of any advantage of doing so.
Last edited:
IMHO, what makes NOS fatigue free is headroom, the intersample overshoot issue mentioned before. And this is a NOS characteristic that is always there no matter the files are processed, compressed or anything.
Intersample overshoot issues are rare in audiophile recordings and if they happen its only a handfull of samples in a file.
The 4 files from Hans from have in total 3 oversample events.
Cant imagine these events ruin the fatigue free feeling for the whole file
Okay, thanks, Marcel. This question is one which I expected this thread to eventually need address as we moved to cover how our findings might impact SDM DACs. If I interpret your points correctly, then it seems that a SDM based DAC, with an bypassed OS FIR interpolation-filter, could also exhibit NOS character. Just as Doede reports for the PCM1794A.
Further to Marcel's SDM exposure, isn't an important reason for the 8Fs up-sampler in front of the SDM to enable the SDM to operate at 64Fs, quite a bit above audio frequencies needed for an effective 40dB noise shaping action as is the case for the 179X.
So removing the 8Fs upsampler and driving the SDM with 8Fs instead of at 64Fs with up-sampler, means that the whole noise shaping envelope already shifts a factor 8 to the left in frequency.
As an example see the image below showing the SDM noise shaped spectrum as measured on a 25R load when Fs is 192Khz and the SDM operates at 64Fs=12.288Mhz on a scale from 0 to 2.0Mhz
Reducing the SDM input frequency to 44.1Khz input in NOS mode, will let it operate at 8Fs=352.8Khz, or at a almost 35 times lower frequency as in the shown recording, shifting the whole picture below a factor 35 in frequency to the left.
See the adjusted frequencies in red in the image, plus the Fs=44.1Khz.
The spectral lines between 60Khz and 200K in the original image are caused by an SMPS feeding the DAC and have nothing to do with the SDM.
In case of Fs=44.1Khz they will shift a factor 35 upwards.
So what I wanted to show is that using no up-sampler in front of the SDM will increase the noise by a significant amount in the audio range.
Hans
.
So removing the 8Fs upsampler and driving the SDM with 8Fs instead of at 64Fs with up-sampler, means that the whole noise shaping envelope already shifts a factor 8 to the left in frequency.
As an example see the image below showing the SDM noise shaped spectrum as measured on a 25R load when Fs is 192Khz and the SDM operates at 64Fs=12.288Mhz on a scale from 0 to 2.0Mhz
Reducing the SDM input frequency to 44.1Khz input in NOS mode, will let it operate at 8Fs=352.8Khz, or at a almost 35 times lower frequency as in the shown recording, shifting the whole picture below a factor 35 in frequency to the left.
See the adjusted frequencies in red in the image, plus the Fs=44.1Khz.
The spectral lines between 60Khz and 200K in the original image are caused by an SMPS feeding the DAC and have nothing to do with the SDM.
In case of Fs=44.1Khz they will shift a factor 35 upwards.
So what I wanted to show is that using no up-sampler in front of the SDM will increase the noise by a significant amount in the audio range.
Hans
.
Attachments
Last edited:
Correct Hans,
as there is no extra 8FS up-sampling and FIR filtering, the noise shifts to the left so to speak
I wrote about this on my website a few years ago.
The positive thing was that as I did not use any digital filtering (FIR), like in the PCM chip or with external digital filters, by using BCK as System clock.
The sound came out as typical NOS. I could compare this with my in those days active DDDAC1543 project, which run over many years and had/has also hundreds of user. Including Jean Hiraga who loved both versions (as it sounded so nicely NOS)
I measured this noise curve depending on FS, see below link
DDDAC 1794 NOS DAC - Non Oversampling DAC with PCM1794 - no digital filter - modular design DIY DAC for high resolution audio 192/24 192kHz 24bit
as there is no extra 8FS up-sampling and FIR filtering, the noise shifts to the left so to speak
I wrote about this on my website a few years ago.
The positive thing was that as I did not use any digital filtering (FIR), like in the PCM chip or with external digital filters, by using BCK as System clock.
The sound came out as typical NOS. I could compare this with my in those days active DDDAC1543 project, which run over many years and had/has also hundreds of user. Including Jean Hiraga who loved both versions (as it sounded so nicely NOS)
I measured this noise curve depending on FS, see below link
DDDAC 1794 NOS DAC - Non Oversampling DAC with PCM1794 - no digital filter - modular design DIY DAC for high resolution audio 192/24 192kHz 24bit
Attachments
I might have missed some posts, but iirc a true NOS-DAC will have a distinct drop off before 20 KHz.
It might depend on program material, but it is usually easily recognisable, and might add a certain character, or flavor to the sound.
In all these home listening tests, how does one compensate for those hf-level differences?
The following is my own, very subjective, take on oversampling, when using HQPlayer to convert pcm to sdm(dsd).
What also is recognisable is that changing the oversampling rates, filters and dither settings, (and for sdm also algorithm settings) leads to significant changes in sq.
This might very well depend heavily on the playback system.
I've experienced, for instance, an acoustic guitar change from sounding as if the strings were way closer to the microphone than the body of it, and vice versa: the strings buried in the lower frequency range of the body.
To add further despair, so to speak, softness, attack and pace as well as depth of imaging can also be tuned by altering these parameters and yes, they do range from "darker", to plain "brighter", and from "fluent" to "artificial" and "deep" to "flat" or "wide".
Of course, my process is different and not all of these parameters might weigh similar in your setup, but it is a description of what these settings can do and how to get to them.
I can't help but think that, together with differing bck/wdck/sck parameters that come into play when switching from 44k1 to higher rates, this might just add to the confusion that seems to go around here.
In my experience, that's perfectly normal.
Iow: mixing true nos-dac owners' perceived SQ with testimonies of non true nos-dac owners might lead to confusion.
Program material (originally high resolution recording down mixed to 44k1) might lead to a different result than true 44k1 material. And which recordings are true 44k1's even? Because this way, settings are mixed from studio to one's home brew up&downsampling settings.
Original classical music might be perceived as responding different to operations than rock/pop music, partly because leading and trailing edge differ. I might be wrong here, but what else could it be?
I can everybody wholeheartedly recommend downloading a trial of HQPlayer desktop and go alter the settings and see what they do. I will be very happy if e.g. all true nos dac owners settle for similar settings. At least, this way, one can play with all the settings and be a bit less dependent on external, out-house processing. Rest assured, I am in no way affiliated with the developer Jussi, or the software. But it's time to bring out some bigger guns;-)
It might depend on program material, but it is usually easily recognisable, and might add a certain character, or flavor to the sound.
In all these home listening tests, how does one compensate for those hf-level differences?
The following is my own, very subjective, take on oversampling, when using HQPlayer to convert pcm to sdm(dsd).
What also is recognisable is that changing the oversampling rates, filters and dither settings, (and for sdm also algorithm settings) leads to significant changes in sq.
This might very well depend heavily on the playback system.
I've experienced, for instance, an acoustic guitar change from sounding as if the strings were way closer to the microphone than the body of it, and vice versa: the strings buried in the lower frequency range of the body.
To add further despair, so to speak, softness, attack and pace as well as depth of imaging can also be tuned by altering these parameters and yes, they do range from "darker", to plain "brighter", and from "fluent" to "artificial" and "deep" to "flat" or "wide".
Of course, my process is different and not all of these parameters might weigh similar in your setup, but it is a description of what these settings can do and how to get to them.
I can't help but think that, together with differing bck/wdck/sck parameters that come into play when switching from 44k1 to higher rates, this might just add to the confusion that seems to go around here.
In my experience, that's perfectly normal.
Iow: mixing true nos-dac owners' perceived SQ with testimonies of non true nos-dac owners might lead to confusion.
Program material (originally high resolution recording down mixed to 44k1) might lead to a different result than true 44k1 material. And which recordings are true 44k1's even? Because this way, settings are mixed from studio to one's home brew up&downsampling settings.
Original classical music might be perceived as responding different to operations than rock/pop music, partly because leading and trailing edge differ. I might be wrong here, but what else could it be?
I can everybody wholeheartedly recommend downloading a trial of HQPlayer desktop and go alter the settings and see what they do. I will be very happy if e.g. all true nos dac owners settle for similar settings. At least, this way, one can play with all the settings and be a bit less dependent on external, out-house processing. Rest assured, I am in no way affiliated with the developer Jussi, or the software. But it's time to bring out some bigger guns;-)
Hi, mterbekke,
The response droop @20KHz, which is due to the sample/hold (zero'th-order-hold) operation of most audio DACs, is not a differentiating factor in our first digital filter test. This is becasue the test file choices being compared were all 44.1KHz rate, and all suffer exactly the same NOS response droop at 20KHz. So, the response droop is not in any way a distinguishing factor between them.
It is a factor in our 88.2 experiment because those files are down -0.75dB @ 20KHz. Since the 44.1 files are down -3.17dB @ 20KHz played NOS, the 88.2 files are raised +2.42dB @ 20KHz in relative comparison. The solution is to EQ all playback to be flat @ 20KHz, however, that would require everyone to uniformly possess an NOS DAC with that capability. There are simply some practical realities we have to try and live with as hobbyists conducting an internationally dispersed investigation.
I don't doubt any of that, and that those elements have the potential to play some role in upsampled file evaluations. The PGGB software, for example, features a number of interesting customizing/tuning options. However, that potential is eliminated in our experiment by the fact that we are all listening to exactly the same test files. Any difference in subjective opinion on them will not be due to the possibility that different listeners are hearing test files produced by different upsampling option settings.
As it turns out, exactly, what constitutes a "true NOS DAC" is something which we a currently working through. The issue being, of course, whether SDM DACs could be fairly considered to offer NOS operation. Up until this point, we were making an attempt to keep SDM DACs out of the investigation only until near the conclusion to reduce the number a investigation variables. However, some of our thread contributors, such DDDAC, are quite adamant that SDM can produce the same subjective NOS sound as multi-bit NOS DACs do. So, there very well not be any confusion added to the results, at all, due to listener reports utilizing SDM based DACs. We shall see.
I'm not certain what you are attempting to express in the above statement.
I believe that those complex concerns are outside the control of our rather simple investigation.
Doede,
So long story short for PCM179X at 44.1Khz in NOS mode:
- yes the SDM noise intrudes quite a bit into the audio range and
- no this SDM noise doesn’t reach a level where it interferes in any audible way with the music reproduction, further confirmed by you with a perfect PGGB 88.2 upsampling where SDM noise is practically shifted out of the audio range that gave no benefit in sound reproduction.
Is this a correct resume ?
Hans
P.S. also confirming that a perfect upsampling does not hurt the signal, as to be expected from a mathematical point of view.
So long story short for PCM179X at 44.1Khz in NOS mode:
- yes the SDM noise intrudes quite a bit into the audio range and
- no this SDM noise doesn’t reach a level where it interferes in any audible way with the music reproduction, further confirmed by you with a perfect PGGB 88.2 upsampling where SDM noise is practically shifted out of the audio range that gave no benefit in sound reproduction.
Is this a correct resume ?
Hans
P.S. also confirming that a perfect upsampling does not hurt the signal, as to be expected from a mathematical point of view.
Last edited:
I consider a true NOS dac chip a chip that eats pcm, digests pcm and, excuse my French, pees pcm.
Anything else needs processing of some sort, usually up/oversampling etc and shouldn't be considered true NOS.
The PCM179x is a hybrid, so even though it's an excellent chip, it's not NOS. Stating that doesn't take away it can't sound or measure splendid, they do, it's just by matter of design it ain't NOS and I can't see this being debatable.
IMHO they are at the center of this.
Oké, let me summarize to get my point across better.
If computations on music can cause detrimental SQ, which is minimized by doing less computations, but computations are done on the files in the studio and at home, then mixing effects of these computations might confuse the results, the best way to go at this is to have more control over the computations at home.
PGGB, iirc, uses only one type of filter, that looks like a sort of Closed Source xxx-taps equivalent in HQPlayer. That's just 1 of the 39 filter options provided. Furthermore, the initial user reports as well as PGGB's own reports were mostly of high praise when ultra high oversampling was used(think from and way above 384KHz). This isn't to say I am against PGGB, but it hardly is the one and only piece of software that should be used if you want to test this hypothesis.
If the above is the case, then studio processing, music type as well as processing at home are the defining parameters and using only PGGB might not get every listener to the point of satisfaction.
My point is not to point out that what you're doing is wrong, on the contrary, it's important to be aware of the cascading effects, so you can better test the hypothesis. The only way to do this is using another piece of software.
It might get rid of the belief that NOS sound, no computations at all, is the only way to get in audio heaven, so to speak.
That's a philosophy lacking meaningful technical specifics. We must translate personal philosophy and subjective hearing in to technical specifics, in order to formulate a practical methodology/solution.
You must guard against the instinct to apply analog process thinking, to assess the appropriateness of digital processes.
We haven't claimed that the PGGB is the only resampler of choice. In fact, our first round of digital filter 'transparency', up/dn 44.1 resampling tests, were processed using Audacity. The author of the PGGB, however, has been graciously supporting our investigation. Plus the PGGB resampler features the best objective specifications we've yet come across, so we've been focused on testing it. In any case, I had always intended that we test/evaluate several differeing resamplers, however, that goal has to be balanced against over taxing everyone's patience to conduct the experiments. I think that one more round of testing to compare one additional resampler would be acceptable. We are certainly interested in finding the most transparent sounding resampling on our quest to identify why NOS sounds characteristically different than does typical OS.
You could greatly assist us in evaluating the resampler you've been suggesting, if you were to download the four 44.1 source files we've using for testing the PGBB resampler, from their Dropbox, and then to up/down 44.1 resample those four source files. We could then conduct one more round of 'transparency' testing utilizing those, with Hans Polak managing the test proceedure.
Is it possible for you to perform the up/dn 44.1 resampling of the source files? I'm asking you this because I presume that you are already in possession of, and familiar with using the resampling tool. Which would releave Hans of needing to acquire and learn to use it, just to create the resampled test files. If you are interested and willing to assist the thread, contact me via PM, and you and I can communicate further about it.
If that test indicates a subjective transparency similar to the PGGB's, I think we can then move to conduct a direct subjective comparison between it and the PGGB, both upsampling to 88.2.
Last edited:
Then also: an SDM dac can be a true NOS dac if the source material was recorded in native SDM format.
What exactly is the meaning of native SDM format.
There is no standard in SDM to start with regarding quantisation levels, number of stages, topology and poles and zeros.
We are trying to find out the reason why a NOS Dac produces a distinct sound, most of the time sounding different compared to a OS Dac.
We are not interested in various recording methodologies but take the recording for granted no matter how it came to life.
That’s why conciously different types of recordings were used in our tests not to base conclusions on one single file.
Hans
There is no standard in SDM to start with regarding quantisation levels, number of stages, topology and poles and zeros.
We are trying to find out the reason why a NOS Dac produces a distinct sound, most of the time sounding different compared to a OS Dac.
We are not interested in various recording methodologies but take the recording for granted no matter how it came to life.
That’s why conciously different types of recordings were used in our tests not to base conclusions on one single file.
Hans
And we are not confusing clipping with interover. Clipping is "all ones" - interover is an "analog" problem in a DA as I see it. One can avoid interovers by introducing attenuation in the digital domain before the actual generation of the analog signal. I.e., one can not look at a digital signal or vaw file and say if it has "interover". It is DAC dependent.
Or?
//