Let's develop an ASR inter-sample test procedure for DACs!

[popej]

Source: https://www.sfu.ca/sonic-studio-webdav/cmns/Handbook Tutorial/Audiology.html

Thanks for reminding this book!
Assuming that music has similar effect like noise, after 15 minutes of listening at 85dB SPL average, we will temporarily lose ability to hear noise at 8dB SPL. Does nature abhor good DACs?

 

[popej]

Do I understand it correctly? You not only get distortion but aliased frequencies too?

 

[signalpath]

Your math for ERB is fine. It is your knowledge of the topic that is the problem. ERB is a simplification of auditory filter bank at one loudness level. It is inappropriate to use as is for audibility of noise. I provide a reference from Stuart in the link I gave you on this. Importantly, if you read the second reference, Dynamic-Range Issues in the Modern Digital Audio Environment, from J. AES, you will see this listening test:

View attachment 323306

As I have highlighted, audibility threshold for such a wideband noise can even go down to negative SPL level of -2 dB. And this is for mono. For stereo the noise adds, lowering the threshold even more.

Also notice how environmental noise did not impact the result. The paper goes into more detail on how directional noise is more audible than diffused noise in a room.

Good paper! Will find and read.

L. Fielder (Dolby) did a lot of work on noise audibility (including perception "under program") a long time ago, like 1970s (?). I'll have to dig up those papers. IIRC, his results were surprising and counterintuitive. In addition to low-level noise, we should also look at basic waveform integrity at very low levels. Distortions / uncertainty in waveform may translate into poor imaging. Such signal aberrations are not really quantifiable by THD numbers at such low levels -- better for FFT analysis. Visual wave analysis is also instructive. For instance, here's a screen shot of one of today's lowest-noise DACs with a 25uVrms sine wave (-106dBFS I think). In our experience, this isn't audible as "distortion" -- but can be audible as image smear. Every acoustic recordist I know will tell you that sound-field accuracy is best heard at the very lowest levels -- just before the reverb tail disappears into the blackness of the room (or, in this case, the electronics). The correlation between very low level waveform distortion and sound-field / image purity is repeatable in blind tests, in our experience.

By the way, we can't accurately view/measure 25uV DAC sine waves directly on our AP2722. At this level, the broadband noise of the AP (-121dBu) starts corrupting the pure sine wave trace. To overcome this, we insert an extreme low-noise preamp at 65dB gain after the DAC (-133dBuEIN), and then calibrate the AP scale back to its proper reading. Below around 15uV or so, the quiescent noise of the 65dB preamp starts corrupting the visible sine trace, which I think would define the dynamic limits of today's best audio test path.

 

[AnalogSteph]

Below are spectrograms of outputs of a few things I have lying around:

• DragonFly Red
• Fiio K3 (USB-C) with and without gain

Both K3 versions have Type C connectors. Is it the older one with AK4452 or the newer one (K3 2021) with ES9038Q2M?

 

[splitting_ears]

Do I understand it correctly? You not only get distortion but aliased frequencies too?

It seems so. But I'd wait for the more knowledgeable people on here to confirm it.
@danadam's graphs are quite puzzling. I'm certainly not used to seeing or interpreting things like this!

 

[amirm]

By the way, we can't accurately view/measure 25uV DAC sine waves directly on our AP2722. At this level, the broadband noise of the AP (-121dBu) starts corrupting the pure sine wave trace. To overcome this, we insert an extreme low-noise preamp at 65dB gain after the DAC (-133dBuEIN), and then calibrate the AP scale back to its proper reading. Below around 15uV or so, the quiescent noise of the 65dB preamp starts corrupting the visible sine trace, which I think would define the dynamic limits of today's best audio test path.

This is definitely an issue. So much so that I put it in my AES paper as DACs clearly have lower noise floor than AP. I suggested they either improve the front-end or build in noise amplifier as you deployed.

 

[danadam]

Do I understand it correctly? You not only get distortion but aliased frequencies too?

Isn't this aliasing a distortion too? Or did you mean harmonic distortion specifically?

AFAIU things, I'd expect aliasing if they oversampled by non-integer factor. Maybe it's some kind of IMD from somewhere?

Anyway, here's how the recorded waveform looks:

And here's another look at frequency domain, a spectrum for each second:

And lastly, peak and rms values for each second:

Spoiler: peak and rms

Recorded Peak  Recorded RMS      Input True Peak
L      R       L      R
-4.22  -4.20   -7.23  -7.21   -  -0.6
-4.12  -4.10   -7.13  -7.11   -  -0.5
-4.02  -4.00   -7.03  -7.01   -  -0.4
-3.92  -3.90   -6.93  -6.91   -  -0.3
-3.82  -3.80   -6.83  -6.81   -  -0.2
-3.72  -3.70   -6.73  -6.71   -  -0.1
-3.62  -3.60   -6.63  -6.61   -  +0
-3.52  -3.50   -6.53  -6.51   -  +0.1
-3.42  -3.40   -6.43  -6.41   -  +0.2
-3.31  -3.30   -6.33  -6.31   -  +0.3
-3.18  -3.17   -6.23  -6.21   -  +0.4
-3.11  -3.09   -6.12  -6.10   -  +0.5
-3.01  -3.00   -6.03  -6.01   -  +0.6
-2.92  -2.90   -5.95  -5.93   -  +0.7
-2.88  -2.87   -5.88  -5.86   -  +0.8
-2.88  -2.88   -5.81  -5.79   -  +0.9
-2.89  -2.88   -5.74  -5.73   -  +1.0
-2.89  -2.89   -5.69  -5.67   -  +1.1
-2.90  -2.89   -5.64  -5.62   -  +1.2
-2.90  -2.90   -5.58  -5.57   -  +1.3
-2.90  -2.89   -5.53  -5.52   -  +1.4
-2.90  -2.89   -5.49  -5.47   -  +1.5
-2.90  -2.89   -5.44  -5.43   -  +1.6
-2.90  -2.89   -5.40  -5.38   -  +1.7
-2.90  -2.89   -5.35  -5.34   -  +1.8
-2.90  -2.89   -5.31  -5.30   -  +1.9
-2.90  -2.89   -5.27  -5.25   -  +2.0
-2.90  -2.89   -5.23  -5.21   -  +2.1
-2.90  -2.89   -5.19  -5.18   -  +2.2
-2.90  -2.89   -5.16  -5.15   -  +2.3
-2.90  -2.90   -5.12  -5.11   -  +2.4
-2.90  -2.90   -5.09  -5.08   -  +2.5
-2.90  -2.90   -5.06  -5.05   -  +2.6
-2.91  -2.90   -5.02  -5.01   -  +2.7
-2.91  -2.90   -4.99  -4.98   -  +2.8

 

[danadam]

Both K3 versions have Type C connectors. Is it the older one with AK4452 or the newer one (K3 2021) with ES9038Q2M?

Right, I must have confused it with E10k. It's K3 with ESS.

 

[KSTR]

This is definitely an issue. So much so that I put it in my AES paper as DACs clearly have lower noise floor than AP. I suggested they either improve the front-end or build in noise amplifier as you deployed.

It is extremely hard if not outright impossible to have lowest possible noise and full protection to 100V++ at the same time in a universal analog front-end. I think this is a problem only the user can solve. If you need high gain and lowest noise, chain in your own favorite low-noise preamp and take care you don't kill its inputs.

 

[KSTR]

I think we are going in circles here.

As of yet, we have no evidence that clipped ISO's are audible compared to non-clipped ISO's with any kind of signal that reasonably resembles music. The steady-state fs/4@45° signal does not need to apply which goes without explanation. And even with this signal the hard-clipping is most certainly inaudible unless you're a bat.

OTOH, other types of ISO distortion, like the infamous wraparound, don't need any further testing as those produce annoying clicks which everybody will readily identify.

What we need is a at least one music track that is "not crushed to death" like someone said but contains quite a few ISO's, and the higher the ISO's the better, of course.
Then I'd happily volunteer to prepare a set of files for blind testing with various clipping thresholds, as outlined in some previous post of mine.

 

[bennetng]

What we need is a at least one music track that is "not crushed to death" like someone said but contains quite a few ISO's, and the higher the ISO's the better, of course.
Then I'd happily volunteer to prepare a set of files for blind testing with various clipping thresholds, as outlined in some previous post of mine.

Correlation between sample rate and audible frequency?

No -- it isn't ultrasonic. I don't understand why this misconception persists; perhaps some people are confusing noise shaping with ringing. Totally different concepts. The ringing is just an effect of taking out >fs/2 frequencies from the signal. They have no frequency components

www.audiosciencereview.com

Like this?

 

[KSTR]

Correlation between sample rate and audible frequency?

No -- it isn't ultrasonic. I don't understand why this misconception persists; perhaps some people are confusing noise shaping with ringing. Totally different concepts. The ringing is just an effect of taking out >fs/2 frequencies from the signal. They have no frequency components

www.audiosciencereview.com

Like this?

Thanks for that, I think this would meet the requirement (ISO's reaching +5dBFS) even though it's pretty synthetic (generated by a general MIDI arrangement?). I'll give it a try...

 

[bennetng]

Thanks for that, I think this would meet the requirement (ISO's reaching +5dBFS) even though it's pretty synthetic (generated by a general MIDI arrangement?). I'll give it a try...

You can scroll up to read the previous posts, it is a track released on a CD.

As for the synthesis, these kinds of sound are generally called Chiptune. I recorded the analog output of my Game Boy in the post below (hardware.flac in the attachment).

FLAC v1.4.x Performance Tests - Page 18

FLAC v1.4.x Performance Tests - Page 18

hydrogenaud.io

Here is a fairly accurate representation of an emulator , rendered the same music using the same sample rate (192kHz).

FLAC v1.4.x Performance Tests - Page 18

FLAC v1.4.x Performance Tests - Page 18

hydrogenaud.io

There are some differences (e.g channel balance) from the hardware recording and of course high noise floor, but I bought the hardware in 1990, so all kinds of aging and defects can happen.

 

[popej]

Isn't this aliasing a distortion too? Or did you mean harmonic distortion specifically?

I have expected anti-aliasing (reconstruction) filter to work. But maybe actually we see distortion caused by clipping in anti-aliasing filter.

 

[danadam]

I have expected anti-aliasing (reconstruction) filter to work. But maybe actually we see distortion caused by clipping in anti-aliasing filter.

Ah! So by "aliases" you meant those big spikes at multiples of 11k and not the small ones at about 5k and 17k? Yes, I think they are the result of clipping in the oversampling filter[*] (which I would call anti-imaging, not anti-aliasing).

4x upsampling not clipped: and clipped:

Late edit:
[*] at least the odd multiples

 

[popej]

Ah! So by "aliases" you meant those big spikes at multiples of 11k

Yes, alias for Fs/4 is the same frequency as its 3-rd harmonic.

4x upsampling not clipped: and clipped:

Looks like a perfect explanation, what is happening in a decent DAC without headroom.

 

[KSTR]

Thanks for that, I think this would meet the requirement (ISO's reaching +5dBFS) even though it's pretty synthetic (generated by a general MIDI arrangement?). I'll give it a try...

Here we go. File is mono content, so only one channel was processed in the following.

Original track, at the largest ISO with more than +5dBFS peak level which is beyond what any DAC I'm aware of would handle without clipping.

Reference track (upsampled to 705.6kHz 32 bit, -6dB gain applied, downsampled to 176.4 16bit):

Clipped (same as above, with hard-clipping at -6dBFS before the final downsampling):

Note: The "dimple on the roof" comes from the downsampling as theory says it should (ringing from the linear phase resampling).

Now let's have some fun comparing the two versions (ABX logs welcomed). I didn't bother as I'm heavily biased anyway towards "no difference" from previous experiences.
I'm aware that this simple test might not fully represent what could happen in practice (notably the analog output stage of DAC might run into saturation with a sticky recovery even when the DAC chip itself could handle +6dBFS peaks properly). But is a starting point at least.

 

[popej]

Now let's have some fun comparing the two versions (ABX logs welcomed).

I did the same and tried. Can't hear the difference. But there is difference in spectrum over 22kHz.
Here 4x oversampled and -6dB version:

Then 4x oversampled, clipped at 0dBFS and -6dB:

 

[KSTR]

But there is difference in spectrum over 22kHz.

Yes, and that's the natural consequence of the clipping.

 

[melowman]

Here we go. File is mono content, so only one channel was processed in the following.

Original track, at the largest ISO with more than +5dBFS peak level which is beyond what any DAC I'm aware of would handle without clipping.
View attachment 323918

Reference track (upsampled to 705.6kHz 32 bit, -6dB gain applied, downsampled to 176.4 16bit):
View attachment 323920

Clipped (same as above, with hard-clipping at -6dBFS before the final downsampling):
View attachment 323921
Note: The "dimple on the roof" comes from the downsampling as theory says it should (ringing from the linear phase resampling).

Now let's have some fun comparing the two versions (ABX logs welcomed). I didn't bother as I'm heavily biased anyway towards "no difference" from previous experiences.
I'm aware that this simple test might not fully represent what could happen in practice (notably the analog output stage of DAC might run into saturation with a sticky recovery even when the DAC chip itself could handle +6dBFS peaks properly). But is a starting point at least.

Here we go. I did 2 passes. First pass scoring 9/10, taking time to get familiar with the tracks; then a second pass scoring 10/10, during which I was faster.

In this instance, it was all about the transient. The reference track has more energy at that peak you're showing in your graph, compared to the clipped track. Subtle but audible.

Meanwhile, I'm trying to cook up the same test but with a more challenging track (at least I think), a track that is already pretty crushed and full of clipping. It should be more difficult (or maybe easier, who knows in advance..).