• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

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

Whether it's audible or not is irrelevant. It's a known characteristic of DAC's and should be tested for regardless.

I'm all for expanding testing where it is important, but this is a non-issue and is solely caused by the content being poorly recorded in the first place. It's not up to D/A converters to "fix" mastering issues, especially at the expense of ultimate performance. What's next? 6/9/12dB headroom? Should the D/A converters fix channel imbalances too?
 
That overload happens in a fraction of a second in some music whereas the distortion products you mention are constant. There is a reason the world is not up in arms about this.

As you know, the fix for this is not free. We lose dynamic range for all content and all times. The real solution would be to avoid this in production of the music, not consumption. Or use DSP volume control upstream of the DAC.

I've already suggested this in this exact thread (https://www.audiosciencereview.com/...le-test-procedure-for-dacs.49050/post-1759469), but why not run a +3 dB intersample over test for only DACs with volume control and only at a -4 dB volume control position?

My experience (also posted in this thread -> https://www.audiosciencereview.com/...le-test-procedure-for-dacs.49050/post-1758216), indicates ESS DACs respond fine if their own volume control is used.

This would help us weed out DACs that have weird behavior without asking for anyone to sacrifice dynamic range.

Michael
 
Last edited:
Whether it's audible or not is irrelevant. It's a known characteristic and anyone testing DAC's these days should include this in their routine. Just for reference, if nothing else.
If audibility is irrelevant then there is no reason to test for it.

Again, this is a mistake in production. A patch that damns the entire music catalog is not the solution.

An elegant one would be for the player to scan your files at import and flag the compensation in the form of metadata. This could also be forced by the user or use an online database. These are the correct, no cost solutions that apply the fix where it is needed and not to every piece of music out there.
 
When I was at AES, I got beaten up by a designer for a major company on why I am testing at 0 dB and not -3. 0 dB is causing detection of clipping in the pipeline of the device. So please don't talk as if I am not doing enough here. I am.
 
When I was at AES, I got beaten up by a designer for a major company on why I am testing at 0 dB and not -3. 0 dB

Considering 0dB testing has been standard (for all the major specifications) since the dawn of digital, he/she had no case.
 
I believe the AK4493 places the digital volume control in front of the interpolator.
FWIW, I've been testing at zero attenuation from the internal volume control.


Take a pure 1kHz sine. At +3dB overload, the waveform and samples look like this:

1727832100923.png


It's very audible.

+3dB overload on complex music you likely won't hear. I've got DAC overload test CD tracks (music) where +9dB is audible, but certainly not unlistenable.
You are not showing intersample-overs here. You show digital hardclipping in the original sample rate.
There will be a tiny amount of intersample overs at the edges in this example waveform
BTW, Audacity cannot show how the reconstructed waveform would (approximately) look like. You have to upsample it yourself.


As you know, the fix for this is not free. We lose dynamic range for all content and all times. The real solution would be to avoid this in production of the music, not consumption. Or use DSP volume control upstream of the DAC.
I'm all for expanding testing where it is important, but this is a non-issue and is solely caused by the content being poorly recorded in the first place. It's not up to D/A converters to "fix" mastering issues, especially at the expense of ultimate performance. What's next? 6/9/12dB headroom? Should the D/A converters fix channel imbalances too?
I'm of the opinion that a DAC should be signal-agnostic and make no assumption if the inputs stream is "ill-formed" or not. Actually it is the reconstruction filter of the DAC which creates the IS-Overs in the first place, therefore it is the place where it should be handled.
It should take care of the issue in a sane way, which would be to provide some headroom for IS-overs like +2...3dBFS and apply gentle softclipping starting at +1dBFS. Internal headroom (before the softclipper) should be +10dBFS. At any rate, broken output signals should be avoided (no funny stuff like wraparound and always keeping the DS modulator in a safe, non-overdriven state).

----:----

As for the testing, I think the truly relevant region for application is the -10 to -30dBFS region (well, right were the ESS hump happens to be). 0dBFS of course must be handled flawlessly but a certain rise of distortion is irrelevant. In other words, a DAC that is worse at 0dBFS than another might still be much better in the relevant region.

And IS-Overs up to a certain reasonable level, say +6dBFS, should be handled gracefully, that is, provide signal fidelity up to about +1dBFS and then start a gentle soft-clipping maxing out at +3dBFS not matter how high the true peak would be. And this should be tested at the most brickwall'ish filter setting as that is the most prone to produce the largest IS-Overs.

And I think the problem is much more relevant in ASRC chips and DSP processing in general because that may corrupt a signal that is reused/recorded etc.
 
Take a pure 1kHz sine. At +3dB overload, the waveform and samples look like this:
Intersample overs are a product of reconstruction filter working above the samples that were not digitally clipped, with level approaching at FS but not necessarily reaching FS.
 
An artificial signal with 22050 Hz tone that might generate +6.6dB peaks during resampling can generated using SoX:
Code:
$ sox -r 44100 -n InterSamplePeak2.wav synth 220501s 0 25 sin 22050 repeat 1 fade h 5 10 5
Source: https://hydrogenaud.io/index.php/topic,98752.msg820149.html#msg820149

Screenshot_2024-10-02_09-16-15.png


That signal can overload anti-aliasing filter during resampling/overampling operation, which can influence the whole audio spectrum.

See some stats on the generated signal:
Code:
$ sox InterSamplePeak2.wav -n stats
Pk lev dB      -0.00
RMS lev dB     -4.26
RMS Pk dB      -0.00
RMS Tr dB     -58.37
Crest factor    1.63
Flat factor     0.00
Pk count           4

Let's try to resample it 8x using "very high" quality SoX preset without adjusting volume. We get 13 clipped samples warning message, and although the 22050Hz signal is be lost, because of the band-width filter, the output file still contains the clipped samples.

Code:
$ sox InterSamplePeak2.wav 8x-v.wav rate -v 352800 stats
sox WARN rate: rate clipped 13 samples; decrease volume?
Pk lev dB       0.00
RMS lev dB    -48.63
RMS Pk dB     -25.70
RMS Tr dB    -199.16
Crest factor  270.10
Flat factor    14.24
Pk count          13

That "clipped 13 samples" warning message disappears when we reduce volume by 6.7dB before resampling operation:
Code:
sox InterSamplePeak2.wav 8x-v-voladj.wav vol -6.7dB rate -v 352800 stats
Pk lev dB      -0.03
RMS lev dB    -53.93
RMS Pk dB     -30.99
RMS Tr dB    -196.63
Crest factor  495.56
Flat factor     0.00
Pk count           3

Alternate way to not generate peaks is to change a phase response of the resampling filter to minimum or intermediate phase. This will not generate clipped warning even without changing the volume.
Code:
$ sox InterSamplePeak2.wav 8x-vI.wav rate -vI 352800 stats
Pk lev dB      -0.87
RMS lev dB    -47.23
RMS Pk dB     -24.34
RMS Tr dB    -198.53
Crest factor  207.99
Flat factor     0.00
Pk count           2
 
I think I agree with all of what was said here. That said, I’m happy to test and report if the OS filter (or SRC/ASRC) has headroom or not, and as per the Nielsen / Lund AES paper.

This is an issue I recently discovered (special thanks to @AnalogSteph) and it can go really nasty as I saw it here.
 
Last edited:
That overload happens in a fraction of a second in some music whereas the distortion products you mention are constant. There is a reason the world is not up in arms about this.

As you know, the fix for this is not free. We lose dynamic range for all content and all times. The real solution would be to avoid this in production of the music, not consumption. Or use DSP volume control upstream of the DAC.
Or - like RME does it in the ADI-2 Pro FS R at least - to let the user decide. This DAC has fixed internal headroom (I think 2.5dB) and additionally lets you activate a 3dB attenuation when using it's ASRC functionality. So each user can chose which metric is more important - distortion free playback or maximum SNR.
 
Intersample overs are a product of reconstruction filter working above the samples that were not digitally clipped, with level approaching at FS but not necessarily reaching FS.

Semantics Pavel. You know full well you need at least 3dB above the loudest possible signal you will encounter in digital to ensure the recording system doesn't hard clip.
 
Actually it is the reconstruction filter of the DAC which creates the IS-Overs in the first place, therefore it is the place where it should be handled.

This is chicken and egg stuff.

Why is there a problem in the very first place? Because the original recorded levels were too high. No other reason.
 
Why is there a problem in the very first place? Because the original recorded levels were too high. No other reason.
No. The DAC has to handle that because it is presented an arbitrary sample stream (might no even be audio/music) and it has to handle that properly (within reasonable limits for which I gave values. No one will expect that +10dB and more IS-overs should be reproduced unclipped).

Back in the old days, with analog reconstruction filters only, the issue existed as well and of course circuit designer were using circuits with enough voltage headroom to handle the low-pass filter's overshoot, like in any other analog signal processing/aquisition chain (same for anti-aliasing filters for ADC).
Why would that need to change if the filter now is mostly digital domain?
 
If audibility is irrelevant then there is no reason to test for it.

Again, this is a mistake in production. A patch that damns the entire music catalog is not the solution.

An elegant one would be for the player to scan your files at import and flag the compensation in the form of metadata. This could also be forced by the user or use an online database. These are the correct, no cost solutions that apply the fix where it is needed and not to every piece of music out there.
While I agree that the root cause for this issue lies in wrong mastering (too little digital headroom), we have to accept that there is a lot of music around with this defect. Not taking care and just blaming mastering engineers does not help at all and does not improve my listening experience ;-)

Even if ripped to a NAS or streamed, this music will require attenuation before digital processing.
This attenuation can be either done downstream via PC (if you are using one), via your streaming device (if it allows for proper volume control) or in your DAC, prior to any SRC, DSP and DA conversion.
A DAC with such a functionality covers all sources, therefore is the most elegant and to be preferred solution IMO.
 
Last edited:
Obviously, the way an implementer should handle the issue is reduce the volume digitally prior to the DAC chip (unless it has a proper internal volume control), like by 3dB, or even better 6.02...dB (0.5x) to avoid the need of redithering (at least with modern DAC chips).

Parallel two DAC chips and get the original noise levels and even lower full scale distortion if you're into the numbers game.
 
Why is there a problem in the very first place? Because the original recorded levels were too high.
Clipping is abuse of media while IS-O is misuse. In the end all it comes to DAC and how does it cope matters. While not to measure it? There is not many features, that really differentiate DACs. This one could be important, since we all notice problems on the recordings side.
 
If audibility is irrelevant then there is no reason to test for it.

Again, this is a mistake in production. A patch that damns the entire music catalog is not the solution.
  1. This is not a mistake in production, it is a mistake in the playback hardware.
  2. We didn't have this problem until oversampled DACs were invented. At worst, the analog output stage of a non-oversampled DAC will flat-top clip the peak. Most early DACs actually had the headroom to reproduce the peaks without clipping. None had DSP filters that could overload and spread the clip event before and after the actual peak. Reconstruction filters were analog.
  3. The peak amplitude of any waveform represented in PCM will exceed the amplitude of the samples unless a given sample happens to fall exactly on a peak. This is a fundamental property of PCM encoding. The playback hardware should be capable of reproducing any reasonable band-limited non-clipped waveform that has been encoded in PCM. 3 dB of headroom is sufficient.
  4. Most CD-format recordings have inter-sample peaks that reach +1 to +3 dBFS. This problem is not confined to a few poorly produced recordings. +1 to +3 dBFS waveform peaks are what should reasonably be expected when individual samples reach the maximum PCM codes.
  5. We cannot go back and remaster 3 million CDs or 46 million songs because we have a problem with our playback hardware.
  6. This is very easy to fix in hardware. It cannot easily be fixed in mastering (especially since the music is already produced).
  7. When an intersample over occurs, the FIR filters in the upsampler and/or reconstruction filter can overload. This overload spreads the clip event over all of the filter taps. This means that the clip event can be spread out by the FIR filter. Some of the distortion will occur before the peak and some will occur after the peak. With a 64-tap FIR filter, the distortion starts 32 samples before the peak and ends 32 samples after the peak. If the filter is running at the 44.1 kHz sample rate, this clip event may span 1.45 ms. If this filter is operating at an 8x oversampled rate, this clip event may span 0.18 ms. In either case, these clip events have a relatively long duration.
  8. Intersample overs may occur many times per second, even on a well recorded track. For example, we found 1129 in the 5-minitue long Steely Dan Gaslighting Abbie track. This is about 4 per second and this is not unusual, nor is this track an extreme example. This track is fairly typical.
  9. Some DACs invert the audio when an intersample peak occurs. The inversion can extend before and after the actual peak (due to the length of the FIR filter) The polarity inversion is caused by a failure to recognize an arithmetic overflow in the DSP. DACs that have an inversion problem, produce very nasty distortion. It is very audible. I believe some of these have been tested on ASR, but this problem was missed.
  10. A digital volume control on a DAC may or may not eliminate the intersample clipping problem when turned down. If an ASRC, an SRC, or an interpolator precede the digital volume control, then the distortion will not be eliminated when the volume is turned down.
  11. An analog volume control will not eliminate the clipping.
  12. Lossy compression systems (such as MP3) tend to create an abundance of intersample overs. A DAC that clips intersample overs will sound especially bad when playing MP3s. MP3 compression can add intersample overs even when none were in the original non-compressed track.
  13. ASR ranks DACs and power amplifiers by SINAD to levels that are purely academic. At the upper end of the chart, the distortion and noise will be lower than 0 dB SPL in the listening room. In other words, the distortion and noise is absolutely inaudible because it is below the threshold of hearing. I couldn't be heard even if the music was not playing. This means that a significant portion of the ASR rankings may represent measurable but inaudible differences, but as readers, we appreciate this information.
  14. DACs that have a wrap-around inversion problem need to be identified to ASR readers. Likewise, DACs that clip no matter where the volume control is set, should be identified to ASR readers. DACs that extend the clip before and/or after the actual peak need to be identified to ASR readers. These are design defects that can cause very audible clipping artifacts.
  15. I personally discovered this issue when doing an ABX test on an Analog Devices AD1896 ASRC. We were using the Steely Dan test track described above, and we could clearly hear the insertion of the ASRC. The ASRC has a measured THD+N of about -133 dB, so the insertion should not have been noticeable, but it was. We also discovered that we could not hear the ASRC insertion if the digital level was reduced by 3 dB before entering the ASRC. This remained a mystery for several years, but it was solved when a customer sent me a +3.01 dBFS 11.25 kHz test tone. This revealed the intersample clipping problem in the ASRC.
  16. The intersample peak clipping problem may be the single most audible defect in modern PCM DACs and DSP devices. I would not write it off as an "inaudible defect that only impacts poorly mastered recordings".
  17. ASR measures THD, noise, IMD, jitter-induced distortion, and other defects down to ridiculously low (inaudible) levels, but as readers, we appreciate this detail. Meanwhile, intersample clipping is the elephant in the room. The short-term THD and IMD produced by intersample clipping dwarfs the other sources of distortion (which are also short-term when playing music).
 
Last edited:
Back
Top Bottom