Can you tell me if my understanding of the "upsampling or not upsampling" controversy is correct?

[Infinite100]

Hello knowledgeable people.

I got my Audeze Maxwell, which atm I use in Bluetooth with the Alternative A2DP Driver, so I have LDAC.
I choose this over the supposedly (according to Audeze) preferable USB Dongle because the latter only supports 48000 and 96000, making the use of exclusive mode in Foobar complicated.
With the Alternative A2DP driver I can select all 4 sample rates, 44100, 4800, 88200 and 96000, so that the only case when I can't use exclusive mode is when I play 192000 files, in which case Foobar will automatically downsample to 96000.

So far so good.

But from what I understand there is another kind of resampling, mostly upsampling, that some people are fond of, = upsampling everything to 192000 or higher, to allegedly improve the sound.

In a comment on this website I've read that "Converting a digital (sampled) signal to a continuous analogue waveform requires interpolation to produce the values between sample points. Doing part of this interpolation digitally (upsampling) simplifies the analogue circuitry and gives better results. That's all there is to it. Whether software upsampling is audibly superior to that built into DAC chips is debatable.".
Now, please correct me if I am wrong, but from the mention of "digital to analog" I understand that the alleged benefit of upsampling is only when using cabled headphones/speakers, not when using Bluetooth or other wireless transmission (like the Mawwell's USB Dongle).
And I also guess that when I decide to be fancy and I use the Maxwell with an Enoaudio Mogami cable and the iFi GO Link (which has a ES9219MQ/Q Quad DAC+ with 32bit HyperStream III) the alleged benefits of upsampling will be rather questionable.
Is this correct?
In other words, I can just forget all this topic and let other people worry about how significant the advantage of upsampling might be in other cases?

Thanks

 

[AnalogSteph]

BT typically has bigger fish to fry than DAC digital filter performance. LDAC for one isn't even actually lossless (unlike what the name would suggest). Fundamentally though, if the DAC in your BT device just happens to have a particularly mediocre digital filter, decent software upsampling could definitely help it out.

These DACs with particularly mediocre digital filters do definitely exist, usually in "utility-grade" converters. No idea what's inside the Maxwell, an ES9219 would have at least one filter option that I'd consider unproblematic and a second that's probably fine as well.

 

[SIY]

to allegedly improve the sound.

That is an excellent choice of word.

 

[Infinite100]

LDAC for one isn't even actually lossless (unlike what the name would suggest).

No idea what's inside the Maxwell, an ES9219 would have at least one filter option that I'd consider unproblematic and a second that's probably fine as well.

It's not the topic of this thread but allow me to ask:
Is FLAC truly lossless?
And if it is not, would its lossyness match that of LDAC?
I explain my (maybe wrong) thought process: let's suppose that FLAC is 95% lossless, and LDAC too.
Does that mean that playing FLAC through LDAC doesn't take anything out of those FLAC? Or would LDAC lose 5% of them?

Another off topic question: supposing that the USB Dongle is lossless, would it be more optimal to have lossless USB Dongle in shared mode (= with forced resample of anything that's not 48000 or 96000), or slightly lossy FLAC in exclusive mode (= no forced resampling)?

Now back on topic, I can ask and report what DAC do the Maxwell use. I doubt though that Audeze would use a mediocre DAC and/or filter.

About the ES9219MQ/Q, from your answer I induce (albeit without being able to understand what it means) that this DAC offers different options of filters and that the choice of which to use falls on the OEM, in this case iFi?
In other words I should ask them what filters they use?
If so, can you help me formulating the question to them? I'd have otherwise no idea how.

 

[SIY]

Is FLAC truly lossless?

Yes.

 

[sam_adams]

. . . upsampling everything to 192000 or higher, to allegedly improve the sound.

It doesn't improve anything since, in most cases, there isn't anything to improve.

Rip from an audio cd to 16/44.1:

Nothing there above 21.5 KHz.

Same track upsampled to 24/192 FLAC:

More bits in the up-and-down part, some noise down around -112 dB at 21.5 KHz—which you can't hear anyway, but nothing above 22 KHz.

File size for the 16/44 AIFF, 50 MB. File size for the 24/192 FLAC, 160 MB—most of which is just useless zeroes.

Open Goldberg Variations mastered at 24/96:

While there is non-zero data above 22 KHZ, it's just noise that you can't hear.

 

[AnalogSteph]

Another off topic question: supposing that the USB Dongle is lossless, would it be more optimal to have lossless USB Dongle in shared mode (= with forced resample of anything that's not 48000 or 96000), or slightly lossy FLAC in exclusive mode (= no forced resampling)?

If you are aware of and working around its limitations like the -0.1dBFS limiter and if decent resampling is being used (even Windows' own is fine these days), there's really not much wrong with shared mode output. I have no qualms using it, but I generally also have plenty of headroom in my digital playback chain.

Now back on topic, I can ask and report what DAC do the Maxwell use. I doubt though that Audeze would use a mediocre DAC and/or filter.

Well, not on purpose of course, but anything battery-operated wireless headphone tends to be very power-constrained and as such there are a multitude of other factors to consider as well.

About the ES9219MQ/Q, from your answer I induce (albeit without being able to understand what it means) that this DAC offers different options of filters and that the choice of which to use falls on the OEM, in this case iFi?

Indeed, see chip datasheet. Be glad it's only two choices (a linear phase fast rolloff and a hybrid fast rolloff, both of which seem sensible enough), the fancier ESS DACs have 8 to offer.

While there is non-zero data above 22 KHZ, it's just noise that you can't hear.

That's not the point when it comes to the potential benefits of upsampling. Better combined digital filter performance is:

https://www.nanophon.com/audio/antialia.pdf

The one case of audibly problematic digital filter performance I've encountered in recent times (I'm sure there were more among crummy old soundcards of the late 1990s) had an FIR filter with ±0.05 dB of periodic passband ripple and a ~3 kHz periodicity, which is decidedly mediocre by modern standards.

 

[sam_adams]

That's not the point when it comes to the potential benefits of upsampling. Better combined digital filter performance is:
https://www.nanophon.com/audio/antialia.pdf
The one case of audibly problematic digital filter performance I've encountered in recent times (I'm sure there were more among crummy old soundcards of the late 1990s) had an FIR filter with ±0.05 dB of periodic passband ripple and a ~3 kHz periodicity, which is decidedly mediocre by modern standards.

That OGV track is not up-sampled, but mastered at that sampling rate. We've seen in most of the recent reviews of DACs that the issues raised by Dunn twenty six years ago seems to have have been largely solved with modern DACs and ADCs. Additionally, the average untrained listener—one who is unfamiliar with what pre-ringing and aliasing artifacts actually sound like—probably would not recognize the effects cited in Dunn's paper—even if they were audible with today's equipment.

Considering that most up-sampled tracks come from CD masters, any issues with the filters in the DACs and ADCs used to create and master the CD, are already baked into the track itself and are part of the sound. Up-sampling the track to a higher sample rate will not remove those artifacts—but it might introduce new ones. The problem with the thinking that up-sampling makes everything 'sound' better, is the same "more is better" approach that seems to be applied to just about everything in the present time.

 

[voodooless]

So let’s consider the question about LDAC and upsampling from the perspective of the codec. LDAC using subband coding, meaning that it splits the audio in various frequency bands, and then applied different compression schemes and (psychoacoustic) modeling on them, giving nitrate priority based on how audible artifacts would be. Meaning that any content above 10 kHz is already somewhat butchered, and anything above 20 kHz is almost pure fantasy

So if you pass 44.1 kHz audio upsampled to say 96 kHz, the codec doesn’t need to spend any bits of the content above 20 kHz. So from that perspective, it’s probably not a waste of (too many) bits.

But is it useful? I think many already answered that question: it’s more likely than not not useful.

It's not the topic of this thread but allow me to ask:
Is FLAC truly lossless?
And if it is not, would its lossyness match that of LDAC?

FLAC is 100% lossless, what goes is, comes out exactly. If you zip and unzip your Word document, do the words change?

LDAC is as lossy as any other lossy codec. The major advantage it proposes to have it it’s higher bandwidth, which most of the time, you probably cannot benefit from anyway, because the 2.4 GHz band it too crowded.

 

[MaxwellsEq]

There are people who believe in upsampling.

 

[Blumlein 88]

That OGV track is not up-sampled, but mastered at that sampling rate. We've seen in most of the recent reviews of DACs that the issues raised by Dunn twenty six years ago seems to have have been largely solved with modern DACs and ADCs. Additionally, the average untrained listener—one who is unfamiliar with what pre-ringing and aliasing artifacts actually sound like—probably would not recognize the effects cited in Dunn's paper—even if they were audible with today's equipment.

Considering that most up-sampled tracks come from CD masters, any issues with the filters in the DACs and ADCs used to create and master the CD, are already baked into the track itself and are part of the sound. Up-sampling the track to a higher sample rate will not remove those artifacts—but it might introduce new ones. The problem with the thinking that up-sampling makes everything 'sound' better, is the same "more is better" approach that seems to be applied to just about everything in the present time.

This reminds me of how strange subjective evaluation works in the minds of those evaluating. On the one hand we have the most direct approach with the idea of purism. NOS R2R DACs are better because of this thinking. And at the same time we have the old time distrust of digital that it is inherently flawed especially at the nasty old 16 bit CD low sampling rate. So the more is better approach where upsampling is beneficial to making that silk purse you always wanted out of that sow's ear.

Obviously even without delving into whether either mode of thinking is true these two are contradictory. With digital there is much of this in the audio public. The ultimate expression being that Sigma-Delta is bad because it isn't 16 or 24 real bits, but instead a few bits heavily over-sampled and digitally filtered. The purist solution being 1 bit delta-sigma heavily, heavily over-sampled and btw the more over-sampled the better as the solution. Not 256, not 512, at least 1024 and hopefully 2048 and more. In fact the more the better without limit with the power of the one true bit.

 

[kyuu]

"Converting a digital (sampled) signal to a continuous analogue waveform requires interpolation to produce the values between sample points."

This is not my understanding of how digital-to-analog conversion works. Thanks to Nyquist, as long as you're sampling at twice the rate as the highest frequency you want to capture, there's no interpolation going on. There's only one analog waveform that is a correct match for the samples, and it's identical to the one that was originally captured by the analog-to-digital conversion. Higher sampling rates, or upconversion, only results in capturing higher frequencies, not greater fidelity of lower frequencies. Though in the case of upsampling it would be less capturing and more inventing.

Someone correct me if I'm wrong here.

 

[SIY]

Someone correct me if I'm wrong here.

You're 100% correct. Under Shannon-Nyquist, a bandlimited analog signal (to 1/2 the sampling frequency) is precisely reconstructed. No interpolation needed or wanted.

 

[oleg87]

Software oversampling would be a good idea for playing back CD quality content on some technically perverse non-oversampling, filterless DAC. Otherwise, any (real) difference is going to be inaudible to utterly insignificant, depending on how much upper treble you can still hear at your age.

 

[AnalogSteph]

We've seen in most of the recent reviews of DACs that the issues raised by Dunn twenty six years ago seems to have have been largely solved with modern DACs and ADCs.

Well, they were pretty much solved by the mid-2000s and we have arguably regressed somewhat in the last decade, though the bottom end saw some improvements still. The adoption of (partially) IIR filters has generally been beneficial in reconciling latency and filter performance demands.

You still see the occasional black sheep every once in a while (in my case it was an IDT 92HD93 HDA sound chip from 2012, ±0.05 dB / 0.3 ms), and more commonly grey sheep (like the CX31993 from the Dell dongle Amir tested recently, ±0.02 dB / 0.3 ms) or light grey sheep (PCM510x, ±0.09 dB / 0.4 ms). The classic manufacturers of audio converters generally try not to embarrass themselves too much. ESS has some filters with highish ripple specs but if you see the actual response these tend to turn out a lot less problematic than the numbers might suggest.

Considering that most up-sampled tracks come from CD masters, any issues with the filters in the DACs and ADCs used to create and master the CD, are already baked into the track itself and are part of the sound.

That is fundamentally correct. That being said, don't underestimate pro-level ADCs, even relatively old ones.

• The Crystal CS5326/27 from 1989 and succeeding CS5328/29 sported a periodic ripple spec of ±0.0005 dB, as did the (1991ish) AD1879,
• for the 1993 AK/CS5390 it was ±0.005 dB (the 5389 had to make do with ±0.01 dB),
• then ±0.001 dB for the 1993 PCM/DF1760 and the 1996 AK5391 (as well as its successor AK5392, the 96 kHz capable AK5393, midrange AK5383 and the 192 kHz capable AK5394(A)),
• ±0.005 dB for the 1996/7 CS5394 and (96 kHz capable) CS5396, and same for the AK4520A (1997), AK4522/3 (1998), AK4524 (1999) and AK4528 (2000) codecs and succeeding AK4620A/B (2005), the PCM1804 (2001), AK5385A/B (2003/5) and PCM4202/4 (2003).
• The PCM4220/2 (2006) even had ±0.00015 dB to offer, or ±0.001 dB with the low latency filter.
• The odd ones out are the 2000s Cirrus chips with ±0.0175 dB / 0.06 ms (@44.1) which I think have at least partially IIR filters and as such are not directly comparable (CS5361/4/6/8, CS5381, CS4271/2). Even as FIR the relatively closely-spaced pre-echo would make them fairly uncritical.

These should cover a lot of recordings from the last 3 decades and change. Older equipment generally used non-oversampling ADCs with analog (IIR) antialias filters, making the discussion of periodic ripple rather moot.

Also, recording at 44.1 kHz straight hasn't been too common for a good while now, it's usually at least 48 kHz or even 96 kHz, occasionally even higher. I guess that's what AKM were banking on when they gave their current (2016) lineup a normal sharp filter with ±0.03 dB / 0.3 ms (@44.1) and only 85 dB ultimate for single and double speed modes, which may have been a tad too optimistic... well, at least you can still use the short delay (IIR) filter of same specs that should be transparent, assuming the equipment manufacturer is merciful enough.

 

[UprightCitizen]

You're 100% correct. Under Shannon-Nyquist, a bandlimited analog signal (to 1/2 the sampling frequency) is precisely reconstructed. No interpolation needed or wanted.

No. The exact work you reference itself use the Shannon interpolation formula so...

 

[BDWoody]

No. The exact work you reference itself use the Shannon interpolation formula so...

So, you don't understand the sampling theorem in other words...

 

[SIY]

No. The exact work you reference itself use the Shannon interpolation formula so...

Shannon Nyquist is not interpolation. It is exact reconstruction.

 

[solderdude]

There are people who believe in upsampling.

There actually is a condition where upsampling is truly beneficial.
When you have a DAC with a poor reconstruction filter (you know the NOS and 'slow' filters) that roll-off in the audible band when used with 44.1/48kHz files you can actually get better results using those DACs when one upsamples to 176.4/192kHz and plays that on such a DAC.
A good upsampling algo will have a sharp filter (that the NOS/slow roll-off filter guys try to avoid) and thus will not show any roll-off nor imaging above 22kHz so will sound correct on those 'flawed' DACs... just as good as any decent (with a proper reconstruction filter) DAC would sound.

That would be a valid reason to upsample and would increase signal fidelity with such flawed DACs.
Those 'NOS DAC believers' probably will probably still believe their 'special DAC' will still be special and will most likely not know the 'steep and horrible sounding pre- and post-ringing' is right there again (because of the filter used in the upsampler) but think the 'magic' of the DAC itself is still present.

 

[kyuu]

There actually is a condition where upsampling is truly beneficial.
When you have a DAC with a poor reconstruction filter (you know the NOS and 'slow' filters) that roll-off in the audible band when used with 44.1/48kHz files you can actually get better results using those DACs when one upsamples to 176.4/192kHz and plays that on such a DAC.

Sure sure. Seems like the better way to go about it is to not use a DAC with a garbage filter in the first place though.