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

[morillon]

Loudness war - Wikipedia

en.m.wikipedia.org

seems in any case assumed in the production of essentially electronic (or mixed) music, or even let's say..."commercial"...

in the end not very demanding rating rendered..(and it is an edifying proof...the other considerations of "hq" resolutions etc. theoretical quality in the field of digital hardware etc. seem a bit "out of place" in this case...)


Sorry

 

[JSmith]

It is wrong.

One would want to be rather cautious suggesting @pkane is "wrong" in this field.


JSmith

 

[tccalvin]

One would want to be rather cautious suggesting @pkane is "wrong" in this field.


JSmith

None of us is perfect and we can all be wrong, so we should all strive to be correct (or correct-ed) with our claims.

If the claim is that the oversampling factor (2x, 4x, 8x...) doesn't matter in the detection of ISPs, then it is not correct. Higher oversampling factors can in fact reveal more of a given peak and they can also introduce clipping where there previously was none and I've shown how.

 

[pkane]

Of course you can. That's what True Peak metering is for. True Peak metering uses oversampling to "fill in the gaps" in the sample stream so that you can know exactly where ISPs are.

The tough part would be predicting where ISPs are going to be before digitizing a signal. With a predictable signal, I guess you could time the sampling so that the individual samples end up exactly where you want them to be.

What I mean is, in a lab, you can replicate results with 100% consistency.



That value is the result of a test done with a perfectly consistent periodic signal like a sine wave, not with actual audio material. If we only ever listened to sine waves, then sure, 4x would be fine. In fact, we could just calculate the exact value of peaks and we wouldn't need True Peak metering at all.

From the Recommendation ITU-R BS.1770-5 document:

“For continuous pure tones it is easy to demonstrate, for example, a 3 dB under-read for an unfortunately-phased tone at a quarter of the sampling frequency. The under-read for a tone at half the sampling frequency could be almost infinite; however most digital audio signals do not contain significant energy at this frequency (because it is largely excluded by anti-aliasing filters at the point of D/A conversion and because ‘real’ sounds are not usually dominated by continuous high frequencies).”

We're talking edge cases, but cases nonetheless.

Yes, sure, the calculation is for a sine wave. But it's conservative, using worst-case sinewave at Fs/2 frequency that is at a level above 0dBFS. Maybe some bat music?

 

[NTK]

Poor wording choice on my part. When I say that a sample "goes over 0dBFS" I simply mean that it's a clipped sample. The analog wave would have to go past that value to be properly reconstructed, but it just can't.



I'm saying that the claim that clipped ISPs (over 0dBFS) are created when recording/sampling a signal is wrong. That's basically only true if you record at clipping levels, which doesn't happen unless we're talking about people that don't know how recording works.



It is wrong. every time you increase the upsampling factor (meaning, you insert samples between samples) the Peak of the digital wave increases in amplitude.

With just 4 samples per wave, we are perfectly fine here:

View attachment 477654

With 9 samples per wave, we're clipping:

View attachment 477655

Let's say you adjust the volume so that you don't clip anymore. You'll be fine until you increase the oversampling factor, at which point you'll be clipping again.

That's just how oversampling works.

No. The internal processing is done using floating point numbers (often in double precision) that have no problem dealing with values > 1.0.

It's the mindless conversion back to integers without checking for overflows that caused the problem you showed.

 

[Mike123]

No. The internal processing is done using floating point numbers (often in double precision) that have no problem dealing with values > 1.0.

It's the mindless conversion back to integers without checking for overflows that caused the problem you showed.

And upsampling isn't done by adding identical samples, but by adding zeros. This example doesn't really correspond to reality.

 

[tccalvin]

Lots of confusion here.

Let's all make sure we understand how oversampling filters work:

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

Nov 15, 2024

it appears then that the filter's job isn't so much "low pass filter" as "reconstruction".

The intent is reconstruction; what it actually does is low-pass filtering. If we start with a sampled signal:

A zero-order-hold DAC could produce an analog signal like this:

This process creates mirror images >20 kHz. Filtering them out reconstructs the signal (with the roll-off inherent to ZOH):
(Ignoring how feasible it is to have such steep analog filter in the real world)

And the principle is the same with oversampling too. At the risk of...

• danadam

• 

Thanks to @danadam for the explanation.

 

[pkane]

None of us is perfect and we can all be wrong, so we should all strive to be correct (or correct-ed) with our claims.

If the claim is that the oversampling factor (2x, 4x, 8x...) doesn't matter in the detection of ISPs, then it is not correct. Higher oversampling factors can in fact reveal more of a given peak and they can also introduce clipping where there previously was none and I've shown how.

I already pointed out that there's an upper limit to the maximum difference between ISPs with oversampling greater than 4x, and it's well less than 1dB, and a whole lot less than 1dB for any real music.

So, let's test your claim with something real. Here's the Steely Dan track (Two Against Nature, Gaslighting Abbie) used in the Benchmark paper. The paper says the highest intersample over measures +0.8 dBFS for this track. My measurement shows +0.9 dBFS.

Here's the comparison. Ref column is for 4x oversampling, Comp is for 64x oversampling.
Note that True Peak value of 0.9dBFS shows no difference between these two oversampling rates:

Let's find a music track that demonstrates a greater difference.

For reference, here's lufs.org processing of the same track:

 

[markstein]

Let's find a music track that demonstrates a greater difference.

For reference, here's lufs.org processing of the same track:
View attachment 477872

Cool site, did not know that.

Here's a track from the CD "E.S.T - Live in Hamburg":

 

[markstein]

Btw - lots and lots of other similar examples in my music library (mostly ripped CDs).

 

[tccalvin]

It's going to be tough to find many commercial songs that have a lot of energy near 20k.

Empirical proof using white noise:

 

[morillon]

It's going to be tough to find many commercial songs that have a lot of energy near 20k.

Empirical proof using white noise:

View attachment 477877

You can t found some.
the use of white noise without practical interest
;-)

 

[pkane]

It's going to be tough to find many commercial songs that have a lot of energy near 20k.

Empirical proof using white noise:

View attachment 477877

Try pink, at least that's a bit closer to spectral content of real music and will not destroy your tweeters Make sure it's bandwidth limited to 1/2 Fs, or else it's not a valid test signal.

 

[tccalvin]

You can t found some.
the use of white noise without practical interest
;-)

It was never my argument that there are lots of commercial songs exhibiting crazy ISPs that cannot be caught using 4x oversampling.

My argument is that as far as audio reproduction is concerned, 4x True Peak metering helps, but it doesn't get rid of the issue. I think I've proven my point.

 

[markstein]

Let's find a music track that demonstrates a greater difference.

For reference, here's lufs.org processing of the same track:
View attachment 477872

SHD: Internal digital headroom?

Hi devteam, Could you please confirm which internal digital headroom to assume for a SHD unit using the latest DSP firmware (40bit filters)? Signal path is: 0dBFS input signal (labeled Dirac01 / Dirac 02 in DC) -> Dirac correction filter -> Matrix Mixer (e.g. Dirac01+Dirac02 for a mono...

www.minidsp.com

Find a screenshot from analyzing my music library using jriver MC (per my knowledge using 4x upsampling) as posted in the minidsp user forum (link above).

Goes up to +4,3dBTP (Gravenhurst - Fires in distant buildings - 2005). +3,7dBTP (Depeche Mode - Exciter - 2001). +3,6 dBTP (Jeff Babko, Simon Phillips - Vantage Point - 2000). Just some examples, you see the list goes on.

This level is computed by 4x upsampling.

 

[morillon]

It was never my argument that there are lots of commercial songs exhibiting crazy ISPs that cannot be caught using 4x oversampling.

My argument is that as far as audio reproduction is concerned, 4x True Peak metering helps, but it doesn't get rid of the issue. I think I've proven my point.

I was just talking about high-frequency levels in audio...nothing comparable to white noise in music... It seems that, if not openly claimed, these loudness war approaches seem to be assumed for music that is actually not very demanding...and therefore...it seems this will continue, indeed it hardly seems to bother...the questions are therefore what methods can be used at the consumer level to moderate things for those who wish to do so?
;-)

It's factual... assumed and well known... and doesn't seem to bother me so much...
Like mp3 128k etc etc

 

[pkane]

SHD: Internal digital headroom?

Hi devteam, Could you please confirm which internal digital headroom to assume for a SHD unit using the latest DSP firmware (40bit filters)? Signal path is: 0dBFS input signal (labeled Dirac01 / Dirac 02 in DC) -> Dirac correction filter -> Matrix Mixer (e.g. Dirac01+Dirac02 for a mono...

www.minidsp.com

Find a screenshot from analyzing part of my music library using jriver MC (per my knowledge using 4x upsampling) as posted in the minidsp user forum (link above). Lots of titles with > 3dBTP. I remember some titles even had >7dBTP.

Share the names of the worst offending albums/titles and I can run some tests on these. The discussion is not that ISOs can't exceed 0dBFS by a lot, it's rather that measuring these using 4x oversampling isn't going to underestimate the peaks by any significant amount compared to larger oversampling rates.

 

[markstein]

Share the names of the worst offending albums/titles and I can run some tests on these. The discussion is not that ISOs can't exceed 0dBFS by a lot, it's rather that measuring these using 4x oversampling isn't going to underestimate the peaks by any significant amount compared to larger oversampling rates.

Hi pkane, I have edited my post and included titles of some of the worst titles.

Edit: Intention of my posts is indeed to show that intersample peaks can be higher than 3-4dB on commercially available albums, so mastering to -1dBTP is not sufficient. All peaks are calculated by 4x upsampling.

 

[Mike123]

Share the names of the worst offending albums/titles and I can run some tests on these.

I mentioned it on the two page before last. "Infected Mushroom Wanted To". Camilla DSP clipped with -4.5 dB attenuate.

 

[tccalvin]

the questions are therefore what methods can be used at the consumer level to moderate things for those who wish to do so?

There are solutions. Read up on previous posts in the thread. You need to lower the volume digitally before the audio hits the oversampling stage (or any kinda filtering). If you're streaming, set normalization to quiet.

The real point we should all bring up is that ISPs aren't tested for in DAC reviews. This means we have no idea if the ISP issue can be solved with any given converter.

So when you ask "how can I fix this?" we can't really give a solution for your specific DAC unless you have the means to test it yourself.

So start asking for ISP testing.

Intention of my posts is indeed to show that intersample peaks can be higher than 3-4dB on commercially available albums, so mastering to -1dBTP is not sufficient. All peaks are calculated by 4x upsampling.

There might be some confusion here because we're talking about measured peaks overshooting by 3-4dB (relative) and songs being mastered with peaks at -1dBTP (absolute).

Mastering at -1dBTP with 4x oversampling (meaning that the loudest sample measures -1dBTP) should be fine for most commercial songs (if people actually did it).