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

[markstein]

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).

I would agree to that.

 

[pkane]

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.

I think you misunderstand. Mastering to -1dBTP means that the mastered file will not exceed -1dBFS when oversampled by 4x. For a track that measures +4dBTP (samples reach +4dBFS when oversampled by 4) the level would need to be lowered by at least -5dB to get to -1dBTP.

 

[EJ3]

Hi there,

It's been a while since I thought about this, but why aren't DACs tested for their ability to handle inter-sample overs (or ISP, inter-sample peaks) in ASR reviews?

Here are some established facts:

1. ISPs are very real, and almost any material that's been released after the 90's can have a significant amount of them 'baked-in' regardless of engineering, artist calibre and technical excellence. Most of the time they stay below +3dBFS, but it's possible to find commercial releases that almost reach +6dBFS. Please also note that there is no mathematical maximum to ISPs, as demonstrated here.
2. ISPs are a necessary by-product of PCM encoding, which means that the 'true peak' values will always be greater than the sample values. This is due to the nature of encoding itself: samples are not the signal but an intermediate representation of it. It becomes a "real" signal only after decoding, ie. reconstruction/interpolation.
3. DACs handle these overshoots differently: some of them distort, and some of them implement an internal safety margin, effectively reducing SNR. For instance, Benchmark and RME do this with great merit. In most cases, this margin adds between 2 and 3dB of tolerance for overshoots.
4. Distortion from ISPs only occurs when the DAC is 'pushed' to its maximum output volume. If the unit has a digital volume control, turning it down will usually solve the issue. However, some others can't due to different design choices (fixed output, analogue volume pots, ...)
5. Sample rate conversion can further increase the reconstructed peak levels due to the Gibbs phenomenon. Lossy encoding also creates many overshoots that are even harder to predict.

Ok, so how could we solve this problem to get better fidelity, even with difficult material?

Firstly, by creating a standard test (which would include a test tone + documented procedure) very similar to the J-test for jitter. Currently, there are many inter-sample test files floating around, but none of them are really established, which true peak levels are all over the place. The test procedure itself is also unclear to many people. I'm sure the whole community here could come up with a very robust, yet simple test.

When the test looks robust enough, adding it to the main reviews would not only add an interesting angle not covered elsewhere, but would also push manufacturers to trade some SNR for higher safety margins to avoid potential playback distortion. In other words, put more emphasis on a "clean" most significant bit (msb) than on the least significant bit (lsb). This would be a significant quality improvement, well beyond the magnitude of a 2dB better noise floor.

Today, most properly designed DACs are well above the human hearing threshold, but sometimes fail to provide adequate protection for a necessary by-product of PCM encoding. Only by publishing objective data and reviews can we encourage more manufacturers to solve this problem. ASR has changed many things for the better in the hi-fi/pro audio world, so I'm sure this would be a great next challenge to tackle.

What do you think?

I think "Why bother?"
I guess that you think "Why not make things more complex than they need to be?"

 

[markstein]

I think you misunderstand. Mastering to -1dBTP means that the mastered file will not exceed -1dBFS when oversampled by 4x. For a track that measures +4dBTP (samples reach +4dBFS when oversampled by 4) the level would need to be lowered by at least -5dB to get to -1dBTP.

Yes, you are correct, I did not consider the condition "when upsampled by 4x". In this case -1dBTP should be fine.

 

[danadam]

Empirical proof using white noise:

Proof of what?

That the samples in 8x oversampling can go higher (or clip) than in 4x oversampling? Did anyone doubted that or claimed otherwise?

Of that the samples in 8x oversampling can go higher by more than 0.6 dB over the samples in 4x oversampling (which is what is being discussed, AFAICT)? But then the screenshot doesn't show how much higher they go.

 

[tccalvin]

Proof of what?

I feel like I'm repeating myself so this is the last time:

The claim that there is no difference between 4x and 64x oversampling in the detection/manifestation of clipped samples is not true. Higher oversampling factors can reveal more clipped samples and/or introduce clipped samples where there previously weren't.

That's all. Let's not get stuck on this.

 

[pkane]

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).

Depeche Mode - Exciter - 2001). +3,6 dBTP

Using DeltaWave, Depeche Mode Exciter produced +4.79 dBFS with 4x oversampling, and +4.80 dBFS with 64x oversampling. So, yes, a tiny bit higher peak with greater oversampling. Hardly worth the bother.

"Infected Mushroom Wanted To". Camilla DSP clipped with -4.5 dB attenuate.

+2.48 dBFS with 4x oversampling and +2.49 dBFS with 64x.

 

[tccalvin]

Of that the samples in 8x oversampling can go higher by more than 0.6 dB over the samples in 4x oversampling (which is what is being discussed, AFAICT)?

An interesting read on the topic of proving potentially infinite ISPs:

True Peak Detection

Russell McClellan In the last few years, a number of different countries have passed laws regulating the loudness of audio in television and other broadcast mediums. Surprisingly, loudness is a dif…

techblog.izotope.com

I can't really give a detailed explanation of this proof since I'm bad at math, but others might be able to.

 

[morillon]

In summary: A simple dedicated True Peak software?
(Is it possible to do this in Multitone? ;-) )

A short, precise list of well-known music files ( 3 ou 4.5db) that would allow us to test our efficient DACs? ;-)

 

[theREALdotnet]

An interesting read on the topic of proving potentially infinite ISPs:

True Peak Detection

Russell McClellan In the last few years, a number of different countries have passed laws regulating the loudness of audio in television and other broadcast mediums. Surprisingly, loudness is a dif…

techblog.izotope.com

I can't really give a detailed explanation of this proof since I'm bad at math, but others might be able to.

There is no limit to how high we can make the true peak, if we have enough alternations of

and

.

The number of alternations we can have is limited by the sampling frequency, though, since we must have at least two samples per period.

 

[tccalvin]

The number of alternations we can have is limited by the sampling frequency, though, since we must have at least two samples per period.

No. the sampling frequency of the audio snippet is constant. It's the duration of the snippet that changes.

In summary: A simple dedicated True Peak software?
(Is it possible to do this in Multitone? ;-) )

A short, precise list of well-known music files ( 3 ou 4.5db) that would allow us to test our efficient DACs? ;-)

You might want to check these (and other) earlier posts out:

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

Oct 2, 2024

An elegant one would be for the player to scan your files at import and flag the compensation in the form of metadata.

For those interested, foobar2000 has this feature built-in:

• edechamps

• 

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

Nov 2, 2023

In the attachment there are 44.1 kHz and 48 kHz files with Fs/4 tones which:

• start at -0.5 dBTP,
• in the last 0.1 of each second ramp up by 0.1 dB
• until they reach +2.9 dBTP.

(there are also 4 pulses at the beginning and at the end: for aligning, to "wake-up" the DAC, etc)

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

• DragonFly Red
• Fiio K3 (USB-C) with and without gain
• Cowon Plenue D
• Samsung dongle
• Speaka (AFAIK it's a clone of DragonFly Black)
• Lenovo T480s

The 44.1 kHz file was played on...

• danadam

• 

 

[pkane]

In summary: A simple dedicated True Peak software?
(Is it possible to do this in Multitone? ;-) )

As others pointed out, various other apps already measure true peak from a file, including DeltaWave. Multitone has some ISO-test tones which lets you test a DAC.

 

[krabapple]

I disagree with Archimago's spin on this topic. If we were talking about analog equipment, then sure, who cares if there are peaks at +0.1dB. All that means is that you're going to get a tiny bit more harmonic distortion.

With digital equipment using oversampling, it doesn't matter if the peaks are at +0.1dBFS, +1dBFS, +10dBFS or +100dBFS. It'll all read zero. That means a completely (square-ish) chopped wave at best, and filter overload at worst.

It will mean that that peak is chopped off to appear as a plateau.

What matters to audibility isn't just that it happens, but how dense (how frequent and close together) such events are, in the track.

 

[morillon]

As others pointed out, various other apps already measure true peak from a file, including DeltaWave. Multitone has some ISO-test tones which lets you test a DAC.

if in addition to the tests available....maybe even more reason for the option to exist in your MT
;-)

 

[tmtomh]

It will mean that that peak is chopped off to appear as a plateau.

What matters to audibility isn't just that it happens, but how dense (how frequent and close together) such events are, in the track.

Yes, a good point.

To be clear, I believe that in terms of best practice and robust engineering, the optimal number of digitally clipped samples and clipped analogue peaks is always zero.

With that said, I have always found it interesting that the 1980s 1600 series SONY PCM ADCs used for digitization of analogue tapes for the purposes of CD mastering and production had an adjustable setting related to digital clipping. You could adjust the sensitivity of what would make the clipping lights come on during recording/digitization. It could be set so that a single clipped sample occurred, or only if a certain number in a row occurred. I believe the other options were 8 or 16, but my memory is not precise and the lowest-sensitivity setting could have been as lenient as 32 clipped samples in a row (but I'm pretty sure the max was 16). Regardless, I've always found it interesting because this was the early days and Sony was super-dedicated to CDs being a high-quality product (even if they ended up getting partially foiled by the record labels' tendency not to provide the best-quality or lowest-gen tapes for the digital transfers). So Sony's engineers must have felt that several digitally clipped peaks in a row were not audible, within reason.

Again, I am in favor of zero clipped peaks, from the original recording device through all ADC stages, through digital mixing, mastering, and resampling. I'm only saying that there must certainly be thresholds beyond which we cannot hear digital clipping, and I would assume it wouldn't be too difficult to develop a "safe threshold plus lots of padding just in case" standard, similarly to how we know that there are "safe" noise floor and distortion levels with audio gear.

 

[markstein]

Yes, a good point.

To be clear, I believe that in terms of best practice and robust engineering, the optimal number of digitally clipped samples and clipped analogue peaks is always zero.

With that said, I have always found it interesting that the 1980s 1600 series SONY PCM ADCs used for digitization of analogue tapes for the purposes of CD mastering and production had an adjustable setting related to digital clipping. You could adjust the sensitivity of what would make the clipping lights come on during recording/digitization. It could be set so that a single clipped sample occurred, or only if a certain number in a row occurred. I believe the other options were 8 or 16, but my memory is not precise and the lowest-sensitivity setting could have been as lenient as 32 clipped samples in a row (but I'm pretty sure the max was 16). Regardless, I've always found it interesting because this was the early days and Sony was super-dedicated to CDs being a high-quality product (even if they ended up getting partially foiled by the record labels' tendency not to provide the best-quality or lowest-gen tapes for the digital transfers). So Sony's engineers must have felt that several digitally clipped peaks in a row were not audible, within reason.

Again, I am in favor of zero clipped peaks, from the original recording device through all ADC stages, through digital mixing, mastering, and resampling. I'm only saying that there must certainly be thresholds beyond which we cannot hear digital clipping, and I would assume it wouldn't be too difficult to develop a "safe threshold plus lots of padding just in case" standard, similarly to how we know that there are "safe" noise floor and distortion levels with audio gear.

There is a similar setting in RME Total Mix. The OVR (clipping) indicator can be set for a threshold of 1-10 clipped samples.

I as well prefer zero clipped samples, just because it is technical possible and there is no logical reason not to do it (other than stupid "loudness war" which becomes less relevant with normalization available in most streaming services).

 

[KSTR]

Using DeltaWave, Depeche Mode Exciter produced +4.79 dBFS with 4x oversampling, and +4.80 dBFS with 64x oversampling. So, yes, a tiny bit higher peak with greater oversampling. Hardly worth the bother.

The choice of filter likely plays a bigger role than the oversampling factor once it is greater than 4x. The more the filter is a perfect lin-phase brickwall (sinc), the larger the ISOs become.

 

[pkane]

The choice of filter likely plays a bigger role than the oversampling factor once it is greater than 4x. The more the filter is a perfect lin-phase brickwall (sinc), the larger the ISOs become.

That’s true, but also depends on content. I spent some time experimenting with different filters for TP measurement. The one I’m using is steep and close to the Nyquist frequency.

 

[theREALdotnet]

No. the frequency of the audio snippet is constant. It's the duration of the snippet that changes.

Ah yes, I was misled by the graphs that suggest the frequency doubled (when in fact the time base changed).

In that case, I would probably chalk this up as a curious aliasing effect. Sampling at only peak, trough and null values can lead to weird outcomes. Nyquist's theorem doesn't guarantee that every assortment of samples will lead to a correct reconstruction of the original waveform. Consider for example a set of samples that all fall on the null crossings of a sine wave, You're left with nothing, i.e. catastrophic, uncorrectable aliasing.

 

[LTig]

[..] Sampling at only peak, trough and null values can lead to weird outcomes. Nyquist's theorem doesn't guarantee that every assortment of samples will lead to a correct reconstruction of the original waveform. Consider for example a set of samples that all fall on the null crossings of a sine wave, You're left with nothing, i.e. catastrophic, uncorrectable aliasing.

This is a violation of the Nyquist theorem, as it states that all signals must have frequency componentes below half the sampling frequency. It doesn' say how much below though, as far as I know.