DEM clocks in perfect sync, 18KHz issue
Hi tubee,
Thanks for your reply,
Very good to remember me of DEM oscillator syncing! I just designed an interesting circuit (concept). DEM oscillators run around 200KHz (component tolerances). So I divided BCK by 16 (74HC161) and connected the output (Q3) 176.4KHz to all PIN 16's of the DAC's trough a 100pF polystyrene capacitor , the 470pF capacitor between pin 16 and 17 on ALL DAC's is removed. Now all DEM oscillators run in perfect sync with BCK! I will post a circuit diagram soon.
Hi, poynton,
Thanks for your reply.
Since most people can't hear those high frequencies, it's no problem the D-I DAC produces triangle shaped waves at those high frequencies
. Important however is that ultrasonic interference that covers up details, you can hear, is greatly reduced.
Hi tubee,
Thanks for your reply,
Very good to remember me of DEM oscillator syncing! I just designed an interesting circuit (concept). DEM oscillators run around 200KHz (component tolerances). So I divided BCK by 16 (74HC161) and connected the output (Q3) 176.4KHz to all PIN 16's of the DAC's trough a 100pF polystyrene capacitor , the 470pF capacitor between pin 16 and 17 on ALL DAC's is removed. Now all DEM oscillators run in perfect sync with BCK! I will post a circuit diagram soon.
Hi, poynton,
Thanks for your reply.
Since most people can't hear those high frequencies, it's no problem the D-I DAC produces triangle shaped waves at those high frequencies
. Important however is that ultrasonic interference that covers up details, you can hear, is greatly reduced.Although I like this idea, I'm wondering why it is called a "NOS" DAC. You've created additional samples from the original signal to increase the effective freq response. Isn't that what the whole idea of oversampling is in the first place? It seems to me, ece, you should stick with the "direct interpolation" monicker (ie DI DAC), just so it is clear that it is not strictly NOS.
You cannot derive 18 bit accuracy from 4 DACs fed a 16 bit data stream... the least significant 2 bits undefined... noise.
When you apply linear interpolation... mechanically or in the digital domain you are adding signals that weren't in the original... and this is noise. It will be most apparent at frequencies of fs/4 to fs/2.
You need to study the Nyquist Theorum and Fourier as well... you can't reinvent the math.
If you display 10 cycles of a 20 kHz signal created by your system, un-filtered, you will see the problem clearly. You will see a 20 kHz signal modulated from full to zero at a frequncy of about 2 Khz.
A digital brickwall will not display the same problem. And a digital filter or a chebychev will "ring" at frequencies near fs... this cures the problem you will see with frequencies nearing fs.
Try it and post.
When you apply linear interpolation... mechanically or in the digital domain you are adding signals that weren't in the original... and this is noise. It will be most apparent at frequencies of fs/4 to fs/2.
You need to study the Nyquist Theorum and Fourier as well... you can't reinvent the math.
If you display 10 cycles of a 20 kHz signal created by your system, un-filtered, you will see the problem clearly. You will see a 20 kHz signal modulated from full to zero at a frequncy of about 2 Khz.
A digital brickwall will not display the same problem. And a digital filter or a chebychev will "ring" at frequencies near fs... this cures the problem you will see with frequencies nearing fs.
Try it and post.
poobah, lets discuss this topic please a little bit, I find it very interesting.
Well, it is the sum of four 16 bit numbers that gives you the 18 bit result. I was saying if we were to use one DAC only, it would have to be an 18bit device to accomodate the result. If two consecutive samples differ in LSB, the 3 points in between need two more bits of resolution to make them right.
You are absolutely right about the distorted shapes of higher frequency signals. Is there a way to reconstruct the original shape (not the thin 44.1kHz samples) while avoiding the ringing produced by more complicated oversampling techniques? I do not think so, there is simply no information present. So I think it is always a compromise - either nice sines (oversampling, a lot of "made-up" information) or proper squares without ringing (NOS, no made-up information).
It is obvious the simple linear interpolation does not cure the final shape distorsion, only makes the resultant signal smoother. When I was testing my 12xTDA1543 DAC with no output filter I did not like the rough-edged shape even for 1kHz. I understand the reasons but I had to "smoothen" the steps with a capacitor parallel to the I/V resistor - I found the sound too harsh without it. I do not mind the shape distorsion, it is what I am feeding the DAC with anyway. But I would have prefered to have straight lines between each sample instead of the steps - i.e. some dense linear interpolation.
Well, it is the sum of four 16 bit numbers that gives you the 18 bit result. I was saying if we were to use one DAC only, it would have to be an 18bit device to accomodate the result. If two consecutive samples differ in LSB, the 3 points in between need two more bits of resolution to make them right.
You are absolutely right about the distorted shapes of higher frequency signals. Is there a way to reconstruct the original shape (not the thin 44.1kHz samples) while avoiding the ringing produced by more complicated oversampling techniques? I do not think so, there is simply no information present. So I think it is always a compromise - either nice sines (oversampling, a lot of "made-up" information) or proper squares without ringing (NOS, no made-up information).
It is obvious the simple linear interpolation does not cure the final shape distorsion, only makes the resultant signal smoother. When I was testing my 12xTDA1543 DAC with no output filter I did not like the rough-edged shape even for 1kHz. I understand the reasons but I had to "smoothen" the steps with a capacitor parallel to the I/V resistor - I found the sound too harsh without it. I do not mind the shape distorsion, it is what I am feeding the DAC with anyway. But I would have prefered to have straight lines between each sample instead of the steps - i.e. some dense linear interpolation.
DEM reclocking
Hi -ecdesigns-
Thank's for your interesting contribution!
In case you didn't know about Henk Ten Pierick's "DEM recklocking":
http://www.diyaudio.com/forums/showthread.php?s=&threadid=11949
I've tried it and it's definitely an improvement.
Hi -ecdesigns-
Thank's for your interesting contribution!
In case you didn't know about Henk Ten Pierick's "DEM recklocking":
http://www.diyaudio.com/forums/showthread.php?s=&threadid=11949
I've tried it and it's definitely an improvement.
Let's consider how all this developed:
Focus on the 20 kHz sine reproduction.
In the beginning it was realized that rough edges contained harmonics up into several hundreds of kHx and beyoind. It was also realized that these harmonics were not the best thing to be feeding into an amplifier because they excite parasitic resonances and other nasty things... modulations and oscillations.
The cure for this was simple enough... a multiple order, 4 to 8 poles, filter. In order to achieve the greatest stopband attenuation and passband ripple, a chebychev was the natural choice because you could do it with less capacitors ($$).
Now some manufactureres would use Butterworth or Bessel filters, for bragging rights, but these filters required more stages to accomplish the same thing.
And so it went... in the beginning analog filters were cheaper than digital filters. So the war centered on how many stages of Bessel a manufacturer would put in.
Then digital came along and FIR (I believe) filters were employed. This wasn't done to improve the sound... it was done to maintain the same degree of suppresion of the high harmonics at a lower cost.
Then the wars started... how many times could you oversample. Oversampling never did anything for the sound, it just allowed the manufacturers to use less and less analog filtering and save $$. Because it saved them money they decided to charge you more for it... the higher oversampling that is.
Now, to my knowledge, linear interpolation, was never used... at least it shouldn't have been. Especially between fs/4 and fs/2 linear interpolation introduces error... it produces harmonics and modulants that were never in the original signal. This is at its worst approaching fs/2. It may sound thrilling... some sonic "zing" if you will, but it detrimental to recreating the original signal.
You need only run a suimulation, or an actual measurement, on a 20 kHz signal to see what happens. You could even model this in EXCEL, in a spreadsheet.
Focus on the 20 kHz sine reproduction.
In the beginning it was realized that rough edges contained harmonics up into several hundreds of kHx and beyoind. It was also realized that these harmonics were not the best thing to be feeding into an amplifier because they excite parasitic resonances and other nasty things... modulations and oscillations.
The cure for this was simple enough... a multiple order, 4 to 8 poles, filter. In order to achieve the greatest stopband attenuation and passband ripple, a chebychev was the natural choice because you could do it with less capacitors ($$).
Now some manufactureres would use Butterworth or Bessel filters, for bragging rights, but these filters required more stages to accomplish the same thing.
And so it went... in the beginning analog filters were cheaper than digital filters. So the war centered on how many stages of Bessel a manufacturer would put in.
Then digital came along and FIR (I believe) filters were employed. This wasn't done to improve the sound... it was done to maintain the same degree of suppresion of the high harmonics at a lower cost.
Then the wars started... how many times could you oversample. Oversampling never did anything for the sound, it just allowed the manufacturers to use less and less analog filtering and save $$. Because it saved them money they decided to charge you more for it... the higher oversampling that is.
Now, to my knowledge, linear interpolation, was never used... at least it shouldn't have been. Especially between fs/4 and fs/2 linear interpolation introduces error... it produces harmonics and modulants that were never in the original signal. This is at its worst approaching fs/2. It may sound thrilling... some sonic "zing" if you will, but it detrimental to recreating the original signal.
You need only run a suimulation, or an actual measurement, on a 20 kHz signal to see what happens. You could even model this in EXCEL, in a spreadsheet.
poobah, thanks for the insightful description.
I have seen the signal shapes (1kHz, 10kHz, 20kHz), both the original waves on computer before burning to CD and at the ouput of the NOS no filter DAC. Apart of the steps instead of lines (the way my generator software Audacity linked the neighbourghing samples), they were almost identical. For sure the 20kHz shape had nothing to do with sine.
So what would you suggest, please? Is there a way to preserve nice-edged squares for originaly square signal and sines for originally sine signal, even for higher frequencies? I am afraid there is no such thing.
I have seen the signal shapes (1kHz, 10kHz, 20kHz), both the original waves on computer before burning to CD and at the ouput of the NOS no filter DAC. Apart of the steps instead of lines (the way my generator software Audacity linked the neighbourghing samples), they were almost identical. For sure the 20kHz shape had nothing to do with sine.
So what would you suggest, please? Is there a way to preserve nice-edged squares for originaly square signal and sines for originally sine signal, even for higher frequencies? I am afraid there is no such thing.
DEM sync schematic & oscillograms
Hi, all
As promised, the DEM sync circuit.
Upper oscillogram:
Upper trace, DEM clock
Lower trace, BCLK
Center: the circuit diagram
Lower oscillogram:
Upper trace, DEM clock DAC1
Lower trace, DEM clock DAC2
From first listening impressions sound seems to have improved
Hi, poobah,
Thanks for your reply, I will certainly have a look at it and try to post the oscillogram of the signal you proposed as soon as possible Then we can have a look at it and make improvements if necessary.
Fact remains however, the D-I DAC sounds significantly better then my NOS reference DAC. So the D-I DAC certainly doesn't seem to make things worse compared to the pure NOS-DAC.
Oversampling issue:
Oversampling means to multiply the sample rate by a specified factor, place zero samples between samples (decimating), then filtering out the extensive noise caused by decimating using a brickwall digital interpolation filter. The proces involves high bitrates at the DAC, and the settling time of the DAC used has to be high enough.
The D-I DAC doesn't use any of the above, it runs at a steady 44.1 KHz just as a NOS-DAC does. Furthermore it needs no filtering at the output stage and is phase linear (the commonly used NOS-DAC's always use a filter of some kind at the output, even output tubes can limit the frequency range, and often passive LC filters are added, just have a look at your NOS-DAC diagram).
Correct me if I am wrong but doesn't frequency limitation and LC filtering at the output stage cause (non-linear) phase errors?, and what about the ultrasonic interference of the mirror image that is located very close to the audio spectrum?
Hi, all
As promised, the DEM sync circuit.
Upper oscillogram:
Upper trace, DEM clock
Lower trace, BCLK
Center: the circuit diagram
Lower oscillogram:
Upper trace, DEM clock DAC1
Lower trace, DEM clock DAC2
From first listening impressions sound seems to have improved
Hi, poobah,
Thanks for your reply, I will certainly have a look at it and try to post the oscillogram of the signal you proposed as soon as possible Then we can have a look at it and make improvements if necessary.
Fact remains however, the D-I DAC sounds significantly better then my NOS reference DAC. So the D-I DAC certainly doesn't seem to make things worse compared to the pure NOS-DAC.
Oversampling issue:
Oversampling means to multiply the sample rate by a specified factor, place zero samples between samples (decimating), then filtering out the extensive noise caused by decimating using a brickwall digital interpolation filter. The proces involves high bitrates at the DAC, and the settling time of the DAC used has to be high enough.
The D-I DAC doesn't use any of the above, it runs at a steady 44.1 KHz just as a NOS-DAC does. Furthermore it needs no filtering at the output stage and is phase linear (the commonly used NOS-DAC's always use a filter of some kind at the output, even output tubes can limit the frequency range, and often passive LC filters are added, just have a look at your NOS-DAC diagram).
Correct me if I am wrong but doesn't frequency limitation and LC filtering at the output stage cause (non-linear) phase errors?, and what about the ultrasonic interference of the mirror image that is located very close to the audio spectrum?
Attachments
phofman,
A squarewave should NOT reproduce with square corners. Assuming that any frequencies above 20 kHz have been eliminated, it should show ringing on all corners and the frequency of this ringing should be 18 khz...
If you are interested, i can give you the formula for a 1 kHz sqaure wave that you could plot in EXCEL.
A squarewave should NOT reproduce with square corners. Assuming that any frequencies above 20 kHz have been eliminated, it should show ringing on all corners and the frequency of this ringing should be 18 khz...
If you are interested, i can give you the formula for a 1 kHz sqaure wave that you could plot in EXCEL.
phofman said:
It is impossible to reproduce the original waveform accurately due to the sampling process. Data is always missing.
So it is actually pointless to discuss whether upsampling, oversampling, direct interpolation etc. is better.
After all, what matters is what you hear.
If that means adding a capacitor across the I/V resistor, so be it.
I want to build this design so I can hear it. If I do not like it, I will not listen any more.
Re: DEM sync schematic & oscillograms
Not on its own. As an add on it is still used, its main proponent being Wadia, though they do have a grander name for the process.
Increase the native sample rate considerably, say tenfold but retain the original Nyquist limit. For example for, a 441000hz sample rate use a limit of 22050Hz instead of 220500Hz.
A more realistic approach would be to optimize your digital filter for the waveform in question as per the Meitner IDAT/ BIDAT dacs. Altis may have produced something similar.
One cannot increase the the sample rate by decimating. Zero padding is an interpolating process. Decimation drops samples.
poobah said:
Not on its own. As an add on it is still used, its main proponent being Wadia, though they do have a grander name for the process.
phofman said:
Increase the native sample rate considerably, say tenfold but retain the original Nyquist limit. For example for, a 441000hz sample rate use a limit of 22050Hz instead of 220500Hz.
A more realistic approach would be to optimize your digital filter for the waveform in question as per the Meitner IDAT/ BIDAT dacs. Altis may have produced something similar.
-ecdesigns- said:
One cannot increase the the sample rate by decimating. Zero padding is an interpolating process. Decimation drops samples.
poobah said:Poynton,
What you are saying is that the very same science that created the sampling process should ignored when trying to implement it.
Hope you didn't build any bridges I'll drive over!
Thanks rfbrw... I need some help here!
Not so !
As rfbrw rightly says, the problem is the sampling rate... it was set too low but we have to live with it and all it's failings.
Take a 20kHZ waveform..
Only 2 samples are available to reproduce the original.
No method of NOS, interpolation, over-, up-sampling can determine the original waveform. There is not enough information.
But we do not want to reproduce 20kHz sine or square waves.
we want to listen to music.
The most important piece of measuring equipment is your ears.
As i said before, this thread is becoming huge indeed.
Nice looking oscillographs from Dem reclock Ecdesigns! If all dacs demodulate on exact the same time, the sound problably enhance also.
The logic pcb is kept rather small, good work (the Philips logic is good choice too). It can easy fiddled in a CDP or dac.
Nice looking oscillographs from Dem reclock Ecdesigns! If all dacs demodulate on exact the same time, the sound problably enhance also.
The logic pcb is kept rather small, good work (the Philips logic is good choice too). It can easy fiddled in a CDP or dac.
