Alternative method for measuring distortion

[solderdude]

Well, yes. I thought the IMD hump was where the real basis was found for the rumors/folklore bearing the label "Sabre glare", in the years since that old (and very short) Head-Fi discussion.

What happened was that people were looking for the 'Sabre glare' and then the IMD hump was seen and that was gladly embraced as 'there you see, there is a measurable difference'. Turned out that wasn't a Sabre thing at all but caused by incorrect loading of the DAC chip. Not all DAC's using a Sabre chip had the 'hump' anyway.

The IMD hump is well below any considered audible levels.

 

[abm0]

OK, but back to the topic: I still wonder how much the Df-metric would penalize a device that did have narrow-band glare that was audible. If that was its only problem and it was perfect in every other way, would it cross over into "not transparent" (above -50 dB) or not?

I guess this goes back to the whole discussion on the importance of separate psychoacoustics-based measurements for first establishing what the transparency threshold for the Df-metric should even be.

 

[SIY]

OK, but back to the topic: I still wonder how much the Df-metric would penalize a device that did have narrow-band glare that was audible. If that was its only problem and it was perfect in every other way, would it cross over into "not transparent" (above -50 dB) or not?

Frequency response errors in the midrange are significant, but I haven't seen any DACs have them yet.

 

[Curvature]

OK, but back to the topic: I still wonder how much the Df-metric would penalize a device that did have narrow-band glare that was audible. If that was its only problem and it was perfect in every other way, would it cross over into "not transparent" (above -50 dB) or not?

I guess this goes back to the whole discussion on the importance of separate psychoacoustics-based measurements for first establishing what the transparency threshold for the Df-metric should even be.

Df-metric is an arbitrary measurement with no psychoacoustic basis. Serge has no useful comments on this aspect other than saying the differences between input/output should be as low as possible, i.e., dodging the issue.

 

[Curvature]

This is interesting. The ADC used is the TI PCM4222EVM, which has a best case THD+N of around -90dB.

From the diagram, Serge measured the headphone out of the Dave in a 32 ohm load using 150mV signal. The 1kHz sine produced a Df result of -95dB, while Amir's AP result using a similar 33 ohm load produced THD+N better than -110dB. So at least in this case the Df result is probably limited by the ADC.

There's no way to compare other figures directly since Amir's setup and test battery doesn't match Serge's.

In any case the main weakness of the Df metric is that it tells you something is wrong without revealing what exactly.

 

[solderdude]

It is probably not (but could be) limited to the ADC (drift, accuracy) but is more likely limited by the software that has to compensate for clock drift.
In any case the culprit for this is not entirely clear, as in determined exactly.

I like the attempt that is made and it will say something about signal fidelity, just like the AP says something about signal fidelity but with a better ADC and different methods dedicated to the task.

 

[Curvature]

It is probably not (but could be) limited to the ADC (drift, accuracy) but is more likely limited by the software that has to compensate for clock drift.
In any case the culprit for this is not entirely clear, as in determined exactly.

I like the attempt that is made and it will say something about signal fidelity, just like the AP says something about signal fidelity but with a better ADC and different methods dedicated to the task.

I think in the case of the 1kHz sine, the ADC is more likely, due to noise + distortion, while with more complex signals it is the combination of clock drift and the DAC reconstruction filter.

The project seems abandoned after 2019.

 

[solderdude]

It is basically a form of nulling that includes ADC and DAC properties, phase shifts and clock drift so it is hard to pin 'the edge of the whole measurement system' on a single component or aspect of that component (the ADC) but as you said it could well be. We don't know though and can only suspect.
The only way to find out if it is the ADC is to repeat the exact same test with different ADCs.
If that yields different results then the measurement method by itself is too limited anyway.
The number/measurements may not even be related to perceived sound quality (which was the idea behind the exercise)
To prove that one would have to do controlled listening tests combined wit measurements using a top notch ADC.

 

[j_j]

I'm a little sad that I'll likely never know enough about audio to evaluate your statement in worthwhile detail.

Well, try running a simple distortion test vs. level, both on the audio interface, and then with a power amp in circuit, for example. Look for high frequency noise well below any sensible threshold. That will give you a good example to start with.

Simply use a long tone that wraps (i.e. generate by FFT), and analyze input and output spectra.

 

[j_j]

I think in the case of the 1kHz sine, the ADC is more likely, due to noise + distortion, while with more complex signals it is the combination of clock drift and the DAC reconstruction filter.

The project seems abandoned after 2019.

One thing one does is uses a single device for both ADC and DAC. This does wonders for clock drift issues. Without clock drift issues, indeed, "interesting" things come about.

 

[Curvature]

Well, try running a simple distortion test vs. level, both on the audio interface, and then with a power amp in circuit, for example. Look for high frequency noise well below any sensible threshold. That will give you a good example to start with.

Simply use a long tone that wraps (i.e. generate by FFT), and analyze input and output spectra.

Thank you.

Masking in particular is a really difficult subject. The concept is easy enough, and I can "hear" masking in action all the time around me, but I can't figure out what I will hear of a signal given basic information like spectrum and level, vs. time just by looking at graphs. Understanding the details of the theory well and making accurate predictions is a feat beyond me.

Then there's binaural unmasking. Whole other set of complexities when you're working with two ears rather than one.

 

[j_j]

Thank you.

Masking in particular is a really difficult subject. The concept is easy enough, and I can "hear" masking in action all the time around me, but I can't figure out what I will hear of a signal given basic information like spectrum and level, vs. time just by looking at graphs. Understanding the details of the theory well and making accurate predictions is a feat beyond me.

Then there's binaural unmasking. Whole other set of complexities when you're working with two ears rather than one.

Hmm. Let me see if I can find something.

AES PNW Meeting Report - Masking: What is it and When Does It Happen?

www.aes-media.org

Going back to this may help you understand some of the concepts, if that helps:

AES PNW Meeting Report - Hearing 099

Perhaps start with the 2019 talk, then go to the 2022 talk.