Does a truly transparent ADC actually exist?

[pkane]

Yep, I missed this. Shit happens ... Anyway it would make sense to run the Multitone software on the actual samples used for blind listening, so one can be sure that the levels are properly matched. Since Multitone does not run on my Linux PC I can't verify myself.

I posted the DeltaWave analysis of the two captures. the levels are matched to better than 0.0049dB, so not an issue. Frequency response shows no real differences in the audible range, and very tiny differences in the inaudible. Phase difference is flat.

There's little chance these two might sound different except for the possible intersample overs: one shows 2dBTP the other 0.8dBTP. Who knows how the equipment that Mani is using to compare these handles them.

 

[KSTR]

I posted the DeltaWave analysis of the two captures. the levels are matched to better than 0.0049dB, so not an issue. Frequency response shows no real differences in the audible range, and very tiny differences in the inaudible. Phase difference is flat.

There's little chance these two might sound different except for the possible intersample overs: one shows 2dBTP the other 0.8dBTP. Who knows how the equipment that Mani is using to compare these handles them.

Well, two ADCs could differ in

• gain (easily corrected for after-the-fact, that is, post-process).
• analog and digital filter passband response. Typical on-chip ADC filters often are short linear-phase filters with significant passband ripple, in the time-domain this equates to pre- and post-pulses around the main impulse.
• analog and digital filter stop-band response. The on-chip filters typically allow for some aliasing, response being only ~10dB down (at best) at fs/2.

Intersample-overs from the feeding DAC can be problematic when the DAC properly reproduces them but the ADC has no headroom at that level. The ADC should clip gracefully with slight overvoltage, though... but maybe does not.


Of course the big question is whether these differences could become large enough to be audible in practical situations, long-term relaxed blind-testing... even when DeltaWave shows a very good match that indicates "most likely inaudible differences" as per our common knowledge and experience.
That's hard to tell but I tend to think by now that it is possible, sometimes...

OTOH, when the numbers indicate we're most likely safe then I wouldn't loose any sleep over it.

 

[pkane]

Well, two ADCs could differ in

• gain (easily corrected for after-the-fact, that is, post-process).
• analog and digital filter passband response. Typical on-chip ADC filters often are short linear-phase filters with significant passband ripple, in the time-domain this equates to pre- and post-pulses around the main impulse.
• analog and digital filter stop-band response. The on-chip filters typically allow for some aliasing, response being only ~10dB down (at best) at fs/2.

Intersample-overs from the feeding DAC can be problematic when the DAC properly reproduces them but the ADC has no headroom at that level. The ADC should clip gracefully with slight overvoltage, though... but maybe does not.


Of course the big question is whether these differences could become large enough to be audible in practical situations, long-term relaxed blind-testing... even when DeltaWave shows a very good match that indicates "most likely inaudible differences" as per our common knowledge and experience.
That's hard to tell but I tend to think by now that it is possible, sometimes...

OTOH, when the numbers indicate we're most likely safe then I wouldn't loose any sleep over it.

Sure. The individual numbers are RMS or averages over a large number of samples, so a few excessive errors might get reduced or averaged out in a long recording, even though they may be short but noticeable during playback and represent an obvious tell in a blind test.

That's why it's important to look over the entire recording, and also, change the spectrum analysis setting to Peak Hold, which uses the peak value for each frequency instead of an average. Here's an example of where clipping and ISOs are present in the recording:

 

[mdsimon2]

In the future I plan on testing DACs with adjustable filters and seeing how that changes the result as well as different ADCs. Also interested in seeing how much magnitude correction without phase adjustment improves the null.

Did a bit more testing with an Okto dac8 pro in to a MOTU Ultralite Mk5 ADC, Okto was clocked via SPDIF from the MOTU. Due to limitations of the MOTU I could only run the Okto up to 96 kHz.

Interestingly enough there is almost no difference (especially without EQ) between minimum phase (FRMP) and linear phase (FRLP) DAC filters. I think my interpretation is that the high pass behavior of the ADC dominates and limits how good the null can be. Or it could be that both have similar magnitude response ripple. Also, interesting that there is still a good amount of magnitude response ripple at 96 kHz which means the pkmetric doesn't change between 44 and 96 kHz. Results shown below. With EQ the Okto does better than the MOTU which I attribute to better better noise performance.

DAC	ADC	Filter	Sample Rate (Hz)	DSP Level (dB)	EQ	pkmetric (dBFS)
Okto dac8 pro	MOTU Ultralite Mk5	FRMP	44100​	-6​	No	-63.1​
Okto dac8 pro	MOTU Ultralite Mk5	FRLP	44100​	-6​	No	-63.7​
Okto dac8 pro	MOTU Ultralite Mk5	FRMP	44100​	-6​	Yes	-114.7​
Okto dac8 pro	MOTU Ultralite Mk5	FRLP	44100​	-6​	Yes	-112.7​
Okto dac8 pro	MOTU Ultralite Mk5	FRMP	96000​	-6​	No	-63.5​
Okto dac8 pro	MOTU Ultralite Mk5	FRLP	96000​	-6​	No	-63.2​
Okto dac8 pro	MOTU Ultralite Mk5	FRMP	96000​	-6​	Yes	-112.6​
Okto dac8 pro	MOTU Ultralite Mk5	FRLP	96000​	-6​	Yes	-117.4​

Here are un-EQ'd magnitude and phase response from REW, not bothering to show EQ'd as they are flat.

All of this leads me to the conclusion that the only way to achieve pkmetric better than -100 dB is to have flat magnitude and phase response.

Michael

 

[manisandher]

The differences are really tiny, Mani. Are you saying you can hear the difference between these?

I see your results Paul, and agree that the two files should sound identical. But for whatever reason, they sound subtly different to me. I don't want to belabour the point though - it's certainly not a hill I'm prepared to die on.

I linked them so that the OP could judge for himself.

Mani.