Using Cross Corelation to lower influence of ADC for DAC measurements

[KSTR]

Anyway I think APx will have a bell ringing about
cross correlation and that the implementation is only a matter of
time .

Ideally, one would include the whole signal chain from the DUT downstream into the CC process, that is, effectively two identical channels strips with input stage, notch and makeup gain going into a stereo ADC to fully exploit CC noise reduction. As mentioned earlier, I don't know if APX series has dual notches and gain stages, though. The older range of analyzers don't.

 

[Rja4000]

Cosmos grade 0. 5 volt range. Smsl sux right channel.

CCA
View attachment 388422

CCM
View attachment 388423
CCM ADC noise floor input shorted
View attachment 388425

CCM. ADC 96KHz. Reduced signal to -0.5dB Also moved adc to reduce mains pick up.
View attachment 388426

I understand your plots above, except the third one.

What does that mean ?

 

[Rja4000]

That's simply measuring the ADC noise floor with its input shorted.

How could that work ?

And what figure are you looking at ?

-125.6 dBFS (A) ?
For a class 0, that should rather be -129dBFS (A)

Of course, it can't be -143.3dBFS A either

I really don't get it.

 

[Rja4000]

A test with Virtins Multi Instrument 3.9.11
Source DAC D50 III Mono
997Hz 0dBFS 44.1kHz
ADC RME ADI-2/4 19dBu range
FFT Size 32k, Dolph Chebyshev 200
BW 20-20k Hz
AES-17 1 octave notch

View attachment 387115

Same with REW 5._40 Beta 50

View attachment 387113

As may be seen, both measure the same SNR (0.1dB difference is not significant)

REW is measuring lower THD and therefore gets higher SINAD (more in line with Notch measurementt)
REW also converges much faster.

But it's good to see this implemented in MI still.

I re-tested with version 3.9.11.1 which adds "In phase" Cross correlation averages

Now, with that option, convergence speed and result is similar to REW's "Cross correlation averaging"
(When unchecked, it's similar to REW's CC Mangitude.)

 

[nanook]

How could that work ?

And what figure are you looking at ?

-125.6 dBFS (A) ?
For a class 0, that should rather be -129dBFS (A)

Of course, it can't be -143.3dBFS A either

I really don't get it.

Measuring the ADCs noise floor with cross-correlation straight away is not possible to my opinion. You need to account for the noise reduction = f(#averages).

JamesDyson probably meant the noise floor of the acquisition system (in contrast to the noise floor with the DAC connected) ?

 

[nanook]

Ideally, one would include the whole signal chain from the DUT downstream into the CC process, that is, effectively two identical channels strips with input stage, notch and makeup gain going into a stereo ADC to fully exploit CC noise reduction. As mentioned earlier, I don't know if APX series has dual notches and gain stages, though. The older range of analyzers don't.

This would mean 2 sliding notch filters that match fairly well (phase is my concern in case the notch is deep).
With software and thorough individual calibration this should be feasible.

 

[FrenchFan]

How could that work ?

And what figure are you looking at ?

-125.6 dBFS (A) ?
For a class 0, that should rather be -129dBFS (A)

Of course, it can't be -143.3dBFS A either

I really don't get it.

View attachment 388517

jamesdyson has a grade "0" ADC

The noise curve represents, I think,
the measurement of the noise, with input at "0" in
stereo mode, this corresponds to 123.49dBFS / 125.6 dBFSA

Which corresponds to the noise for the cross correlation
measurement.

In "mono" mode we will certainly see ~ 128 dBFSA

On my iso B I have ~ 122 dBFS ~ 124dBFSA in "stereo"

@jamesdyson, I noticed that by increasing the
sampling frequency we measure a slightly better ADC noise.

Try at 192Khz.

 

[Rja4000]

jamesdyson has a grade "0" ADC

The noise curve represents, I think,
the measurement of the noise, with input at "0" in
stereo mode, this corresponds to 123.49dBFS / 125.6 dBFSA

Which corresponds to the noise for the cross correlation
measurement.

Yes, that's probably it.
Grade 0 should be -129dBFS A with open inputs in mono mode.
So -126 or so in stereo.
Could be lower with input shorted, but I don't know how much.

This has nothing to do with CC, though.

 

[Rja4000]

I meant the cosmos without dac connected and with its xlr inputs shorted. If the noise of the 2 channels is primarily uncorrelated, would it not tend towards 24 bits 144dB?

OK, so what we see us the amount of correlate noise between the 2 channels.
Not sure how usefull that is, though.
Dud you try the same in 32 bits ?

 

[FrenchFan]

This would mean 2 sliding notch filters that match fairly well (phase is my concern in case the notch is deep).
With software and thorough individual calibration this should be feasible.

"WRONG PROPOSITION"

It is possible to make
the measurement of the 2 channels in delayed time.

It can be done with an E1DA adc, it is necessary
to record on the same channel 2 *wav at
different times (in mono)

The ADC noise will be decorrelated in time therefore on the FFT.

Well, it is necessary to double the measurement time, OK

But if you use the E1DA ADC measurement in "mono"
we gain in addition "3db" it will be a plus the day
when the DACs will go below the 135dBFS of THD + N

In addition it will be necessary to adjust only one ADC and only one
notch filter.

We can take by software take alternately on a mono signal
a block of sample and the next one and correlate them, the noise will be
different in time from one sample to another. It will certainly be necessary
to look closely at the possibility of overlap.

Well, I think there are still some ideas that will emerge.

 

[KSTR]

Of course. I didn't mean the adc chips themselves. I meant the cosmos without dac connected and with its xlr inputs shorted. If the noise of the 2 channels is primarily uncorrelated, would it not tend towards 24 bits 144dB?

It would go to -infinity after infinite amount of time if the noise is completely uncorrelated.
In a "quick" test it took about 20k averages to go below -144dBFS (20Hz...20kHz bandwidth).

 

[KSTR]

It is possible to make
the measurement of the 2 channels in delayed time.

But that de-correlates (and thus reduces) the DUT's noise, which typically is not what we want.

 

[KSTR]

This would mean 2 sliding notch filters that match fairly well (phase is my concern in case the notch is deep).

Yes, the lower harmonics might see a phase difference between channels when the notches are not 100% the same. But as far as I have seen even a 10 degree delta does not cause any significant error.

 

[FrenchFan]

Yes, that's probably it.
Grade 0 should be -129dBFS A with open inputs in mono mode.
So -126 or so in stereo.
Could be lower with input shorted, but I don't know how much.

This has nothing to do with CC, though.

If a little bit.

When we have this noise value, which I grant you,
is only an approximation of the noise on a real measurement,
We can with 5*log(N) approximate the noise floor of the ADC

For jamesdyson with 6346 correl, so -19 dB of correlation gain
on his last measurement

We can say that the noise of the ADC = -123.49 -19 = -142.49 dBFS

Like this we know approximately the distance between the noise
of the ADC and that of the measurement.

For jamesdyson he has left if we look at his last measurement
DAC noise (-134dB) - ADC noise (-142.49) = 8.5 dB of noise margin
his measurement really tends towards an increasingly accurate value.

 

[FrenchFan]

But that de-correlates (and thus reduces) the DUT's noise, which typically is not what we want.

Right, there I am bad
Thank to correct

 

[FrenchFan]

Yes, the lower harmonics might see a phase difference between channels when the notches are not 100% the same. But as far as I have seen even a 10 degree delta does not cause any significant error.

Isn't the phase corrected on the notch filter?

 

[nanook]

Yes, the lower harmonics might see a phase difference between channels when the notches are not 100% the same. But as far as I have seen even a 10 degree delta does not cause any significant error.

Using two notches that do not match properly would thus mainly introduce errors in the frequency region of the fundamental (because here the phase walk is pretty steep, particularly for a deeper notch than the moderate 30dB the APU is adjusted to).

The calibration in REW does account for amplitude and phase:
-> The noise of the source gets amplitude and phase aligned for both channels (good resolution in the notch region presumed)
(If I remember correctly, the calibration in REW is common for all channels, so this isn't possible as-is, right?)
The ADC noise is still uncorrelated.

So with an individual calibration of both channels even this should work in principle.
(This basically is what I meant for a possible solution for the AP, just that it would have to cover the frequency range)

 

[Rja4000]

Yes I tried 192, sanes 96kHz. I don't have an explanation for the improvement from 48kHz.

This is documented by Yvan.
The ADC chip has slightly better performance when a bit hot, and increasing sampling frequency makes it heating.
He even sold, at some point. a resistor to heat the chip and regulate its temperature.

(Edit: still sold here)

 

[KSTR]

Is it correct to allow it to go below the fundamental bit depth of the system?

Yes.

 

[FrenchFan]

Doesn't Apx have a measurement without notch filter??

In this case it could make the measurement in 2 passes

One for the precision of the THD (with notch)
The other for the precision of THD+N (without notch)

???