Audibility thresholds of amp and DAC measurements

[Hayabusa]

Sorry, didn't catch what you're suggesting.

REW has this possibility:

Then instead of plotting for instance the second harmonic distortion for 1kHz at 1Khz it will plot it at 2Khz the place where you actually hear that harmonic.

 

[nick_l44.1]

THD seems a lot lower up high.

You missed my comment that the top graph was obtained at -12 dBFS.
Here, an accurate measurement of the CS43131 THD 10 kHz with a notch.
-.5 dBFS

-12 dBFS

The height of h9 and h7 does not stand a chance for low THD.

Why, pray tell, would I care of there was an 80dB down distortion product only at, say, 70kHz? That would not make any sense at all to be a "measurement object".

Good question. I don't know for sure, but sometimes it's the only explanation for what I hear. Since the ultrasound region contains (or may contain) interference products of audible frequencies, any perturbations from extraneous signals or noise in this region can reduce the ability of hearing to distinguish or recognize individual quiet sounds.

You do understand that it's not only harmonics that are measured, yes? Unless distortions and noise above 20kHz (or maybe 25kHz just to be really, really safe) are powerful enough to cause equipment or drivers to do something nonlinear, what do you believe is the problem.

I didn't understand the point of this question. When measuring THD, the energy of the areas corresponding to the harmonic frequencies is summed up without taking into account whether it is tone or just noise.

The IMD test is less clear to me because I haven't had much practice with it and therefore it is difficult to evaluate the result.
Yes, I see the DL200 has a small IMD product at 1kHz. I assume because of the higher THD at 19 and 20 kHz, but find it difficult to correlate it with possible audible effects.
In a few words, I'm trying to measure the difference in sound I'm hearing. I have found differences, but can't draw a line to highlight the more significant ones.

Then instead of plotting for instance the second harmonic distortion for 1kHz at 1Khz it will plot it at 2Khz the place where you actually hear that harmonic.

Do you mean regular single tone fft instead of THD vs frequency plot as I attached above in this post?

 

[Sokel]

You missed my comment that the top graph was obtained at -12 dBFS.
Here, an accurate measurement of the CS43131 THD 10 kHz with a notch.
-.5 dBFS
View attachment 456556
-12 dBFS
View attachment 456557
The height of h9 and h7 does not stand a chance for low THD.

You're right, I did miss that, so did it again at about -12db:

Not much of a difference but for fairness.

Edit: and a quick and dirty 10kHz one so we can see the harmonics up there:

That's what I usually like about the old AKM ones, not SINAD kings by any means but consistent and decent across the board.

 

[nick_l44.1]

As to "loss of transparency" I guess I'll have to ask how you demonstrate that.

As I wrote above, rustle level sounds become indistinguishable from noise. Ideally, transparency is the effect of listening to a live performance, not a recording.

so did it again at about -12db

It's hard to say why the results are so different. I have no reason to doubt the accuracy of my ADC (Cosmos ADCiso + Cosmos APU). Maybe our DACs are different in some way. I tested it on two dongles with bare cs43131 and cs43198 with almost the same success. It seems @jkim also got similar results with jcally jm20 max.

 

[Sokel]

It's hard to say why the results are so different. I have no reason to doubt the accuracy of my ADC (Cosmos ADCiso + Cosmos APU). Maybe our DACs are different in some way. I tested it on two dongles with bare cs43131 and cs43198 with almost the same success. It seems @jkim also got similar results with jcally jm20 max.

Oh, there must be a misunderstanding, it's not cs43131 and cs43198 measured in my plot, just the 20 years old AKM chips of a dirt-cheap interface from back then.

The post was about that in this 20 years some DAC chips are still far from perfect as freqs go up.
That's solely from an engineering point of view, it's not about audibility.

 

[nick_l44.1]

Oh, there must be a misunderstanding, it's not cs43131 and cs43198 measured in my plot, just the 20 years old AKM chips of a dirt-cheap interface from back then.

The post was about that in this 20 years some DAC chips are still far from perfect as freqs go up.
That's solely from an engineering point of view, it's not about audibility.

Oh, I see. There is an opinion that these CS chips are also based on very old designs.

 

[j_j]

Good question. I don't know for sure, but sometimes it's the only explanation for what I hear. Since the ultrasound region contains (or may contain) interference products of audible frequencies, any perturbations from extraneous signals or noise in this region can reduce the ability of hearing to distinguish or recognize individual quiet sounds.

You need to look at much more of the error spectrum. Get a plot of the entire error spectrum inside the audio bandwidth. That's what matters.

I didn't understand the point of this question. When measuring THD, the energy of the areas corresponding to the harmonic frequencies is summed up without taking into account whether it is tone or just noise.

Every distortion measurement I know of is THD+Noise. This means that the test signal is removed, and everything else is measured within a given bandwidth. Therefore, noise resulting from image distortion, IMD, etc, all shows up in the measurement, WITHIN THE AUDIO BANDWIDTH, and that's what matters.

Noise at high frequencies outside the audio bandwidth is less of a concern unless it's huge and going to result in IMD or driver/amplifier problems.

The IMD test is less clear to me because I haven't had much practice with it and therefore it is difficult to evaluate the result.
Yes, I see the DL200 has a small IMD product at 1kHz. I assume because of the higher THD at 19 and 20 kHz, but find it difficult to correlate it with possible audible effects.
In a few words, I'm trying to measure the difference in sound I'm hearing. I have found differences, but can't draw a line to highlight the more significant ones.

You need to start thinking "what that winds up inside the audio bandwidth is the problem here".

Do you mean regular single tone fft instead of THD vs frequency plot as I attached above in this post?

I don't mean "single tone" at all. I mean error spectra. Error signal spectrum. For whatever the input is.

 

[Hayabusa]

Do you mean regular single tone fft instead of THD vs frequency plot as I attached above in this post?

I mean to do this in this kind of plot:

It would shift the effects of higher order harmonics further away from the audible band and would be more realistic on audibility.

 

[nick_l44.1]

It would shift the effects of higher order harmonics further away from the audible band and would be more realistic on audibility.

Ah! Found this option, but it's disabled. I didn't manage to activate it in any way, and couldn't find any information about what it may depend on.

You need to look at much more of the error spectrum. Get a plot of the entire error spectrum inside the audio bandwidth. That's what matters.

Of course it matters. But it's been a long time since I've come across equipment of such poor quality that distortion is strongly reflected in the audio band. Even if we look again at the CCIF IMD test of the DL200, where a 1kHz tone occurs. -114 dB is not much considering that the 19 and 20 kHz tones are unrealistically high. I could be wrong, but it's doubtful that this test unambiguously shows a problem that can be heard.

It would be great to get the error spectrum, but how can this be done?

Every distortion measurement I know of is THD+Noise. This means that the test signal is removed, and everything else is measured within a given bandwidth. Therefore, noise resulting from image distortion, IMD, etc, all shows up in the measurement, WITHIN THE AUDIO BANDWIDTH, and that's what matters.

It is possible to measure THD independently of noise. Ultrasonic noise is not of too much interest, as you wrote and with which I completely agree. That is why I am comparing THD up to 20 kHz over a wide measurement bandwidth. And intermodulation, which leads to tones in the audio range among other things, is also a product of harmonic distortion.
Of course, the thermal and modulated noise levels have to be controlled too, but in my case I didn't find any problems with that.

By the way, do I understand correctly that you don't think recordings with sampling rates above 44/48 kHz can have better quality or is there a nuances?

 

[j_j]

It is possible to measure THD independently of noise. Ultrasonic noise is not of too much interest, as you wrote and with which I completely agree. That is why I am comparing THD up to 20 kHz over a wide measurement bandwidth. And intermodulation, which leads to tones in the audio range among other things, is also a product of harmonic distortion.
Of course, the thermal and modulated noise levels have to be controlled too, but in my case I didn't find any problems with that.

Why do you WANT to measure only at harmonic frequencies? This makes it literally impossible to calculate a THD for a complex signal with a filled spectrum. Simply removing the original when calculating the SNR, that will tell you what you've got in terms of problems, and why it might or might not be a difficulty. What you need to measure is the effects of nonlinearities, harmonic, IMD, or other. This captures noise, modulated noise, gain drift, thermal aspects, the whole kaboodle. So why not measure what actually matters, not some useless subset of the data?

So, since you claim you can measure THD independent of noise, just how do you do that? Sample the spectrum only at harmonic frequencies? With what bandwidth? How do you calculate the harmonic structure of a complex signal? Look at the "buzz tone" there's a script for somewhere in the "general audio" forum. Look at the spectrum. What are the harmonics of that? How do you handle the sum/difference components of all that? Are sum/different components due to nonlinearity count as THD?

By the way, do I understand correctly that you don't think recordings with sampling rates above 44/48 kHz can have better quality or is there a nuances?

Well, that's more complicated than it ought to be. Young persons, very young people, can hear about 20kHz, although there's effectively no content to speak of that gets to your ears at such frequency with any substantial level, and most of that is rejected by the eardrum and the bones that couple sound to the cochlea. I am well aware that if you get close enough to an instrument there are some remarkable harmonics.

Now go look at the attentuation of 20kHz with distance in 40% humidity air, then farther up than 20khz, and tell me when those stop mattering? https://sengpielaudio.com/calculator-air.htm is a very handy tool here. For instance, the the second harmonic (40000 Hz, of a 20000 Hz tone) is attenuated about 1.25 dB per meter of air propagation. How close are you sitting?

However, that does not address the time-domain effects of antialiasing filters that are required to capture at 44.1 or 48 kHz. There is roughly an infinite variety of such filters, both IIR and FIR. This gives you many choices, phase shift below 20kHz (sharp IIR), potential (although unproven) pre-echo nearing 20kHz, or some from Column A and some from Column B. Or you can take the symmetric FIR and move some of the zeros from outside the unit circle to inside, and thus make filter with some pre-echo and some phase shift. And so on.

Does it matter? I doubt it. I do know that if you sample at 64kHz, with a slow-rolloff filter from 25kHz to 32kHz, then the impulse response of the filter is brief enough to not interact with the narrowest cochlear filter impulse response, and thus entirely unlikely to have a problem. This is true for either IIR or symmetric FIR. Period.You can also go to 96khz sampling, like many do, and use a very wide transition band filter, and be safe from even hypothetical problems. OF COURSE most 96kHz filters in ADC's try to be sharp, and "widen the bandwidth", which means you don't actually shorten the impulse response as much as you could. Then people use half-band filters, but that's a different problem.

 

[nick_l44.1]

Why do you WANT to measure only at harmonic frequencies? This makes it literally impossible to calculate a THD for a complex signal with a filled spectrum. Simply removing the original when calculating the SNR, that will tell you what you've got in terms of problems, and why it might or might not be a difficulty. What you need to measure is the effects of nonlinearities, harmonic, IMD, or other. This captures noise, modulated noise, gain drift, thermal aspects, the whole kaboodle. So why not measure what actually matters, not some useless subset of the data?

Because that is the only thing that can be done with high accuracy. Of course, a single tone is the easiest load for a DUT to handle. But since all distortion other than noise is caused by the same nonlinearities, a THD test will certainly show them all. Even trying to measure TD of a normal multitone doesn't give a stable result, so its accuracy is questionable. I admit that the way complex signals are analyzed can show the presence of some problem, but not its “size”. Also complex signals have a possible problem of harmonics overlapping with signal tones. In the worst case, if you build a signal from a harmonic sequence (1k + 2 + 3 + 4 + ... + 20), the result will be hard to notice harmonic distortion at all, only intermodulation ones. And the “buzz tone” is built exactly on this principle.
Ok, I don't see any problem to just try “buzz tone”. This spectrum is obtained on DL200 HPA output. It's a quite dirty, but all distortions are below -120 dB. What do you think about it?

Next, 9039s

This is exactly what is written above. Distortion is as if there is no distortion at all and it is definitely unreal... and beautiful)

Well, that's more complicated than it ought to be.

Thanks for the detailed response! Something to think about.

 

[j_j]

Because that is the only thing that can be done with high accuracy. Of course, a single tone is the easiest load for a DUT to handle. But since all distortion other than noise is caused by the same nonlinearities, a THD test will certainly show them all. Even trying to measure TD of a normal multitone doesn't give a stable result, so its accuracy is questionable. I admit that the way complex signals are analyzed can show the presence of some problem, but not its “size”. Also complex signals have a possible problem of harmonics overlapping with signal tones. In the worst case, if you build a signal from a harmonic sequence (1k + 2 + 3 + 4 + ... + 20), the result will be hard to notice harmonic distortion at all, only intermodulation ones. And the “buzz tone” is built exactly on this principle.
Ok, I don't see any problem to just try “buzz tone”. This spectrum is obtained on DL200 HPA output. It's a quite dirty, but all distortions are below -120 dB. What do you think about it?
View attachment 456816
Next, 9039s
View attachment 456815
This is exactly what is written above. Distortion is as if there is no distortion at all and it is definitely unreal... and beautiful)


Thanks for the detailed response! Something to think about.

The next step is to subtract the input spectrum. The bottom plot looks very, very linear. The top one shows a set of lines at n*500, which is the IMD resulting from the difference in tone frequencies, but compared to some I've seen, it's really not too bad. You did read the info along with the tone, yes? You can analyze it at a length to avoid any need of windowing, yes? It's generated specifically to allow that (continuous across any block length that is an integer number of the initial signal before it's repeated many times).

Oh, and note the buzz tone is NOT harmonic or anything like harmonic. It is anharmonic. There is a reason why the files are called "dontlistentome.wav". It is designed specifically to cause distortions to pile up at 500, 1000, 1500 ...

 

[nick_l44.1]

It is anharmonic

Oh, my fault. That's because I didn't see any harmonic distortion in the 9039s output.

The next step is to subtract the input spectrum. The bottom plot looks very, very linear. The top one shows a set of lines at n*500, which is the IMD resulting from the difference in tone frequencies, but compared to some I've seen, it's really not too bad. You did read the info along with the tone, yes? You can analyze it at a length to avoid any need of windowing, yes? It's generated specifically to allow that (continuous across any block length that is an integer number of the initial signal before it's repeated many times).

I'm not familiar with matlab, and subtraction in rew gives a visible error. So I just summed all IMDs manually and got -115 dB. Or taking into account the scale shift (DAC to ADC) -109.3 dB. That doesn't seem like much. But it is very disconcerting that the test shows for the 9039s a clean range of almost -160 dB, whereas the same CCIF IMD test gives at best -130 dB.

 

[j_j]

Oh, my fault. That's because I didn't see any harmonic distortion in the 9039s output.

I'm not familiar with matlab, and subtraction in rew gives a visible error. So I just summed all IMDs manually and got -115 dB. Or taking into account the scale shift (DAC to ADC) -109.3 dB. That doesn't seem like much. But it is very disconcerting that the test shows for the 9039s a clean range of almost -160 dB, whereas the same CCIF IMD test gives at best -130 dB.

Still quite decent equipment, there. Instead of looking for harmonics with taht signal, look for multiples of 500Hz. The tones are at prime number frequencies as far as the FFT length, with a stride of 499.xxx Hz apart (fixed number of lines in the fft that is used to create the signal. That means that nonlinarities for interaction between tones all comes at that 499.xx multiple, and the lowest order is 499.xx itself. It's very close to 500Hz, but locked to the FFT value of the stride in the signal generator.

 

[nick_l44.1]

Still quite decent equipment, there. Instead of looking for harmonics with taht signal, look for multiples of 500Hz. The tones are at prime number frequencies as far as the FFT length, with a stride of 499.xxx Hz apart (fixed number of lines in the fft that is used to create the signal. That means that nonlinarities for interaction between tones all comes at that 499.xx multiple, and the lowest order is 499.xx itself. It's very close to 500Hz, but locked to the FFT value of the stride in the signal generator.

Yes, the DL200 is very good, although it is possible to notice that its sound is not ideal, even on the line output.
Of course, I have seen the spectrum of the original buzz tone and what the DAC adds during conversion. -109.3 dB is the sum of N*500 Hz tones.

 

[j_j]

Yes, the DL200 is very good, although it is possible to notice that its sound is not ideal, even on the line output.
Of course, I have seen the spectrum of the original buzz tone and what the DAC adds during conversion. -109.3 dB is the sum of N*500 Hz tones.

That's still not terrible at all. In what situations do you actually hear distortion? Note that the buzz signal is very close to a pessimal signal. If you change the limits, you can probe the DUT only with the higher frequencies, or only with the lower frequencies, or your choice. That is, if you have matlab handy, or octave, sorry. By changing the range of the individual lines, while keeping the same set of frequencies (just not including all of them) you can find out where the distortion comes from. I wouldn't faint in shock if it was mostly due to higher frequencies.

 

[nick_l44.1]

That's still not terrible at all. In what situations do you actually hear distortion? Note that the buzz signal is very close to a pessimal signal. If you change the limits, you can probe the DUT only with the higher frequencies, or only with the lower frequencies, or your choice. That is, if you have matlab handy, or octave, sorry. By changing the range of the individual lines, while keeping the same set of frequencies (just not including all of them) you can find out where the distortion comes from. I wouldn't faint in shock if it was mostly due to higher frequencies.

I finally found time to continue the tests. And it turned out that last time I tested DL200 I made some mistake, apparently I didn't notice a bad contact. Today's tests show distortion about 20 dB lower.

As tempting as this idea is, the error levels are too low to even measure due to noise in both the DAC and ADC.

 

[rdenney]

Just a comment that comes to mind reading the last series of posts. I’ve been measuring a Holman preamp using an HP 8903B distortion analyzer. It works by digitally notching out the fundamental of the test tone and then measuring the RMS voltage of everything else in the test bandwidth. It comes with filters to narrow the test bandwidth—mine has a 400-Hz high-pass filter (useful for testing for power-supply noise) and low-pass filters at 30 KHz and 80 KHz. The total bandwidth of the device is 10 Hz to 110 KHz. The part that’s relevant to a comment above is that the distortion measurement changes by only a fraction of a dB when switching between the 30 and 80 KHz low-pass filters for all the devices I’ve tested so far. That tells me that functioning audio equipment generally shows very little effect of distortion products above 30 KHz. An amp that rings at high frequencies because a faulty design might be different, and in that case the fault would be exposed.

The AC voltmeter measures harmonic distortion and noise by integrating across the bandwidth of the machine—it’s not an estimation and it matches within a microvolt the transfer-standard HP 3456a voltmeter that has sensitivity and resolution down to 100 nanovolts (-140 dBV measuring a 1-volt signal).

Pulling just harmonic distortion out of that would simply sum in the frequency domain across the peaks of the fundamental’s multiples, it seems to me. I don’t know how REW does the math. But it doesn’t take much attenuation of the upper harmonics to add very little to the total. They are just applying a bit of negligible fuzz to the waveform.

Rick “fun with logs” Denney

 

[j_j]

I finally found time to continue the tests. And it turned out that last time I tested DL200 I made some mistake, apparently I didn't notice a bad contact. Today's tests show distortion about 20 dB lower.
View attachment 457160
As tempting as this idea is, the error levels are too low to even measure due to noise in both the DAC and ADC.

You still have clear IMD in the distortion spectrum, though. Those peaks at 100, 1500, etc. Interestingly enough 1K is clean, so whatever distortion is very much antisymmetric, I suspect. Still that's low.