Oversampling: Who Does It Best?

[Charles Hansen]

I'm not sure why you have chosen Linn to be your digital hero. Their products are competent at best, but surpassed in performance by many others.

 

Funnily enough, after I talked about the issue of compounded rounding errors, then so did you. Perhaps if you weren't so guided by picking holes in other products/people, you would have got there directly from the 1 page (so clearly marketing, but not misleading) Linn DS upsampling document.

 

I disagree. I think that referenced webpage was misleading. Even more misleading (and embarrassing) was this related webpage:

 

http://docs.linn.co.uk/wiki/index.php/Dither

 

This document points out that the only two processing operations are the oversampling and the volume control. (They neglect to specify how many stages are used in their 8x oversampling filter.) First they admit that the initial release didn't use any dither and that there were low-level artifacts as a result. This is just plain embarrassing. Any kid who took a college level course in DSP would know that dither was required.

 

Then they show the output of the digital volume control with the volume turned down to -40 dB. But this is just a pathetic joke.

 

They pretend that this graph represents the output of the entire device. But it doesn't. They have to reduce this signal to 24 bits to present it to the DAC chip (hopefully they will use dither this time!!). Let me assure you that the graph of the output of the DAC chip would be very different than the graph they show.

 

They even claim that the graph shows the result of truncating to 24 bits, which is insane. A 24-bit signal is not going to have a noise floor of -190 dB! They are playing fast and loose with their computer-generated graphs. This web page is simply designed to mislead. It does a good job of it. If you want to see what the unit really does, check out the measurements of the Klimax DS at the Stereophile website:

 

http://stereophile.com/digitalprocessors/308linn/index.html

 

Here is the Stereophile graph:

 

 

 

This is at full volume. The distortion is from the analog circuitry. Rest assured that new artifacts would pop up with the volume level reduced.

 

Looks a little different than the Linn-supplied graph, no? Here is the one with dithering:

 

 

 

The Linn graph is of an imaginary mathematical world where the noise floor drops 50 dB (!) and there are no distortion artifacts. No wonder they call one of their DACs the "Majik".

 

As for your other webpage talking about their "Up-sampling", there are at least two blatant errors:

 

1) The page states, "The up-sampler takes digital audio at any of the supported sample rates and up-samples it to a constant rate of 352.8kHz or 384kHz. This allows the up-sampling filters in the third party DAC to be bypassed, thus removing a potential source of noise and signal distortion."

 

This is sheer nonsense. In the first place, the DS products have a DAC. Why would someone use a third-party DAC with it? In the second place, what DAC (third-party or otherwise) can accept a 384 kHz signal?

 

2) The page states, "Various filter shapes were tried including linear-phase, minimum-phase, apodizing, etc. The resulting filter has a linear-phase characteristic and is what Linn term a ‘true’ upsampling filter. This means that it is fully attenuating at the Nyquist frequency and so prevents the formation of ultrasonic-images."

 

Again, more nonsense. I am surprised that Linn tried minimum-phase filters and rejected them when every other company that has tried them has adopted them. But to each his own I suppose. The real joke is that what they did adopt is the very definition of an apodizing filter! But they tried to pretend it was something else, I guess.

 

[Miska]

In the second place, what DAC (third-party or otherwise) can accept a 384 kHz signal?

 

Are we now talking about chips or devices? I know a few devices devices capable of that.

 

When it comes to chips, there are bunch of DAC chips that allow 384 or 768 kHz input sampling rates.

 

Third option is custom solution like for example dCS has.

 

 

[ziggyzack]

First they admit that the initial release didn't use any dither and that there were low-level artifacts as a result.

 

Originally, digital volume was only offered in their bottom of the range players. They didn't expect people to use it much, but it was popular, and there was demand from users across the range. So, it was brought up to spec and released across the range (just a firmware update kicked off from my PC, it's all networked after all). How would I know that if they hadn't 'admitted' it? You're still throwing mud. I find them extremely open, that's one of the reasons I like them.

 

As for the dither point, you're wrong again. It would have helped if you'd included both graphs from the Linn document. They show one graph without dither, and one with, to show the effect. There is no suggestion that the graph represents measurements taken from the device outputs. In fact the document says 'To illustrate the effectiveness of the new dithering process, the following two graphs show the frequency spectrum of a sine wave after it has passed through the digital volume control with and without dither. Here's the link again if anyone's interested, it's only one page: http://docs.linn.co.uk/wiki/index.php/Dither. It's tucked away in their technical pages for those who go looking, and any questions are answered on their forum. They're not shouting these figures from the rooftops where they may be taken out of context.

 

And now one of your 'blatant errors':

 

The page states, "The up-sampler takes digital audio at any of the supported sample rates and up-samples it to a constant rate of 352.8kHz or 384kHz. This allows the up-sampling filters in the third party DAC to be bypassed, thus removing a potential source of noise and signal distortion."

 

This is sheer nonsense. In the first place, the DS products have a DAC. Why would someone use a third-party DAC with it? In the second place, what DAC (third-party or otherwise) can accept a 384 kHz signal?

 

Take a breath... they are referring to the DAC chip(s) used in the DS range as third-party DACs, to distinguish them from other DS elements designed in-house. Majik DS uses a single Wolfson WM8740. Akurate DS and Klimax DS both use dual WM8741s. And both of these DACs accept 384/352.8 KHz signals in 8fs mode, which also bypasses the chips' upsampling filters.

 

I don't know anything about the precise nature of their filter, so I can't deal with your blatant error number 2, but on the balance of probabilities I reckon you're talking more bollocks.

 

For what it's worth, here's John Atkinson's conclusion, seeing as you quoted him:

 

This is excellent measured performance. After I had finished testing the Klimax DS, I loaded some of my hi-rez master files on the NAS—specifically the 24-bit, 88.2kHz master files for Attention Screen's Live at Merkin Hall—and gave a listen. Wes had it right: this is one great-sounding component, particularly in its freedom from high-frequency grain and its low-frequency definition. Wow!—John Atkinson

 

ZZ

 

[Audio_ELF]

I'm not sure why you have chosen Linn to be your digital hero. Their products are competent at best, but surpassed in performance by many others.

Is this really an acceptable / appropriate think for a manufacturer to be posting... I'm sure Gilad Tiefenbrun may feel similar about Ayre products!!

 

Eloise

 

[barrows]

The Klimax DS in a relatively familiar system context, and was quite impressed with its sound quality, despite the fact that its output stage uses IC Opamps (which I usually do not like) and the power supply is one of Linn's proprietary switchers. Too bad this unit is ~$20K, which is about $16K beyond what I can consider spending on a digital front end! But for those who can afford it, and take the time to match a system, pretty good sound is certainly possible.

I would love to see what Charlie's company (Ayre) might produce for a DAC at the $**K price level though... I expect such a product might be very impressive, considering what the QB-9 is capable of at under $3K.

 

[ziggyzack]

Lundahl transformers for the Klimax DS, op-amps in the mid-range Akurate DS (a mere £4.5k now, since it's recent facelift).

 

So your no-op-amp preference is in tact (unless it was an ADS you were listening to)

 

ZZ

 

[barrows]

I believe you are mistaken. The Klimax DS has IC opamps, followed by the Lundahl trannys. Check out the 6 Moons review for internal pics. Good performance with transfomers requires that they be fed from a low impedance source, and I doubt the direct output from the Wolfson DAC chip offers a low enough impedance to drive a tranny for low distortion.

Although I do expect the Lundahls are partly responsible for the sound of the Klimax DS (perhaps due to the limited bandwidth of the transformers eliminating RF noise at the output...)

 

[ziggyzack]

I see what you mean. Thanks Barrows.

 

[Miska]

A 24-bit signal is not going to have a noise floor of -190 dB!

 

Mostly depends on length of FFT etc, since the noise is spread over number of bins and more bins there are, lower the per-bin-level.

 

Here's one digital volume example. Left channel originally 0 dBFS and right channel -60 dBFS, both have -60 dBFS gain applied, thus left one now at level -60 dBFS and right one at -120 dBFS. Left channel has flat dither while right channel is employing noise shaping. Data is stored as a 96/24 WAV and shown here.

 

 

 

[PeterSt]

Let me say that it's always guessing a bit how an analyser's manufacturer wants us to treat it, but what I have learned is that it should be "calibrated" against a virtual 0dBFS. So, for example, if my DAC would output 1.5VRMS max, it lacks 3dB of official output level, and thus I'd calibrate the max signal to be 0dB under 0dBFS - which means the noise level will go up with 3dB.

 

This may be somewhat more hard to follow, but when you'd think about the "honesty" in this, it's more clear soon. Thus :

When my DAC does not have sufficient output so to speak - and it is 3dB under the standard (which latter would be 2VRMS), it undoubtedly implies that my amps need to gain more in order to achieve the standard output (net). This within itself obviously means the noise will be gained with that. And so, the analyser does exactly the same when I move the -3dBFS towards -0dBFS.

 

The other way around counts just the same, but needs different thinking;

When my DAC outputs 2.5VRMS this is 3dB too much (according standards), and when it would exhibit a noise level of -144dB the signal sticking out 3dB over official 0dBFS, I can move down the signal by 3dBFS (in-analyser), and the noise will be a -147dB. Indeed, something the analyser officially can't measure, but "can show" (by this means).

This too is honest, because when the DAC is capable of outputting a so high level while the noise stays that low, it just *is* that low - and it is relative to the signal level. So, initially the analyser shows -144 (its max), but the signal sticks out 3dB. Total is -147dB.

 

Notice that by this means you are able to let stick out harmonics from the noise floor, which otherwise would drop with the (illegal ?) surpression of it all - and which is just happening in-analyser. IOW, the Linn picture not showing anything looks like a case of that to me, as does your digital volume example, Miska. So, I guess this will be about your mentioned negative gain, which will also surpress the noise. Maybe that has a purpose, and maybe I understand only half of it all, but I wouldn't know what such a picture could tell me. Ok, apart from the impossibility the noise being that low.

 

All 'n all, just using a familiar figure of my own DAC, at 2uV of inherent noise (end of interlink) this will be just under the limit of the analyser, while it shows -144dB (with indeed a fair amount of depth of the FFT). So, someone may do some math here how much noise it takes to get to -190dB or something, which I can but won't, because I don't think it is ever possible to be that low.

 

As said, I may understand only half of it all, and especially can't know the purpose of such pictures which *may* be there afterall. All I know is that I think I apply it all in a most honest way, or at least in a way which can't make figures better than they really are.

 

Where needed I am happy to stand corrected !

Peter

 

[Miska]

it should be "calibrated" against a virtual 0dBFS

 

Since there's no DAC involved in the figure I posted, just WAV -> processing -> WAV, all in digital, there's nothing to calibrate. 0 dBFS is defined as a signal amplitude reaching maximum and minimum sample values. So there's no VRMS values that would need calibration anywhere. The figure shows results of 60 dB attenuation of originally 0/-60 dBFS signal.

 

So this shows the data that would go to a 24-bit DAC. Just a dump at the point between application and audio interface driver.

 

Ok, apart from the impossibility the noise being that low.

 

There's nothing impossible. Each time you double FFT length, noise floor drops by 6 dB, since the same amount of noise gets distributed over 2x number of frequency bins. So in order to see distortions, you'll need to use long enough FFT to get the noise floor low enough not to hide the harmonics.

 

Also use of averaging lowers the noise floor, since the noise process is supposed to be random. No averaging used in my figure.

 

Second is that with noise shaping you can tailor the frequency distribution of the noise. So bandwidth can also have an impact. Any of the modern DACs being example. They can give you >100 dB audio band SNR with just a few bits. SACD can give roughly 120 dB SNR within 20 Hz - 20 kHz band with just a single bit.

 

My example shows result of very gentle noise shaping, easy to compare the two channels.

 

This was just to show that the digital volume control doesn't have to be the limiting factor, even at 24-bits.

 

Tech specs for the figure, IIRC; 96/24 output data, 65k FFT, Hann-window, no averaging, logarithmic scales.

 

Edit: I've attached the sample file, so everybody can run what ever kind of analysis they want...

 

[Charles Hansen]

Take a breath... they are referring to the DAC chip(s) used in the DS range as third-party DACs, to distinguish them from other DS elements designed in-house. Majik DS uses a single Wolfson WM8740. Akurate DS and Klimax DS both use dual WM8741s. And both of these DACs accept 384/352.8 KHz signals in 8fs mode, which also bypasses the chips' upsampling filters.

 

You are right. I mis-read the text and thought that when they said "third party DACs" that they meant complete DAC boxes from another manufacturer. Since the only manufacturer of complete boxes that also made their own chips was Sony (they sold off their semiconductor division a couple of years ago), that is not the way I would have phrased it, but clearly I jumped to the wrong conclusion. Sorry about that.

 

[Charles Hansen]

Mostly depends on length of FFT etc, since the noise is spread over number of bins and more bins there are, lower the per-bin-level.

 

Yes, I know that. But my point wasn't that you could use an FFT with thousands of bins and reach an arbitrarily low noise floor.

 

Instead my point was that the graphs Linn presented were only of a mathematical abstraction that is not representative of the performance attained at the output of the device.

 

Since we have not yet devised a "headphone" that connects our brains directly to a digital signal, we must rely on the analog output of the DAC box. In my opinion this is what Linn should be showing on their website, as anything else is misleading and deceptive to the average consumer.

 

[Miska]

Since we have not yet devised a "headphone" that connects our brains directly to a digital signal, we must rely on the analog output of the DAC box. In my opinion this is what Linn should be showing on their website, as anything else is misleading and deceptive to the average consumer.

 

OTOH, since I'm not selling any piece of hardware, I cannot represent exact analog domain results either (and I try to avoid listing specs on the website for that reason). All I can tell is what goes out to the conversion hardware from the digital processing. Rest depends on quality of the hardware. And I can give rough estimates on how it will likely perform based on tests with the hardware I have.

 

In any case, the discussion here was about _digital_ volume controls, noise, rounding errors and distortions in _digital_ domain. I mostly wanted to emphasize that if it sounds bad, either the digital implementation is not very good or the DAC has poor low level linearity or other problems in the analog domain. SNR/THD+N with 24-bit data is lower than with any real world analog power amp I've seen. But good if someone has exceeded dynamic range of 24-bits in analog domain. Even though it requires quite some extra effort to reach SNR of best analog pre-amps even with the best DAC chips, and these best pre-amps don't use potentiometers.

 

 

[Charles Hansen]

In any case, the discussion here was about _digital_ volume controls, noise, rounding errors and distortions in _digital_ domain. I mostly wanted to emphasize that if it sounds bad, either the digital implementation is not very good or the DAC has poor low level linearity or other problems in the analog domain. SNR/THD+N with 24-bit data is lower than with any real world analog power amp I've seen.

 

In theory this all makes sense. But the real world is not always the same as theory.

 

For example, when we were working on our digital filter we could feed it 24-bit data from a DVD-Audio disc and were using 32-bit coefficients. (They sounded better than either 16-bit or 24-bit coefficients, even when using a CD as the source.) So the output was 56 bits. We ended up using a 64-bit accumulator to avoid any overflows.

 

Then we needed to reduce this back down to 24 bits to feed to the DAC chip. There was no point to truncate, but we listened to rounding and two different types of dithering. As all of these methods only affect the 24th bit going into the DAC chip, I was not expecting to hear any difference whatsoever. In the first place the DAC chip is not linear to anything close to 24 bits. In the second place, the noise floor of anybody's analog circuitry is well above that.

 

Yet in blind testing I could easily pick out each of the methods. I have no good explanation for this except to note that we can hear far more things than we can measure. After my experiences with the design of our digital filter, I would never throw away resolution by using a digital volume control.

 

[Gregor Samsa]

Mr. Hansen,

 

Allow me to express my appreciation that, in your early emulation of Linn, you did not adopt their product naming convention. Something called a "Klimax Kontrol" should only be sold out of the back pages of Hustler.

 

Gregor

 

[Miska]

For example, when we were working on our digital filter we could feed it 24-bit data from a DVD-Audio disc and were using 32-bit coefficients. (They sounded better than either 16-bit or 24-bit coefficients, even when using a CD as the source.) So the output was 56 bits. We ended up using a 64-bit accumulator to avoid any overflows.

 

Maybe the difference is here, since I don't use fixed point arithmetics, I never liked it. Everything is 64-bit floating point (11-bit exponent and 53-bit mantissa including sign) until final stage. Exponent representation is especially nice for multiplications. On Linux I could also easily switch to using x86's extended precision (15-bit exponent and 65-bit mantissa including sign), but I haven't seen any benefits from the increase. Rounding to nearest is performed by hardware (level of that is 2^-53).

 

Yet in blind testing I could easily pick out each of the methods. I have no good explanation for this except to note that we can hear far more things than we can measure.

 

This is why I test these algorithms by purposefully emulating extreme attenuations of 120, 108 and 96 dB. I do this by transforming music files into 192/4, 192/6 and 192/8 for listening. With a good algorithm, it sounds exactly the same, just the noise floor becomes more audible. And this noise shouldn't sound bad in itself.

 

After this, if there is still audible degradation with real attenuation using DACs, I know the problem is in the DAC electronics and should be fixed there.

 

After my experiences with the design of our digital filter, I would never throw away resolution by using a digital volume control.

 

Now the interesting part is that for example the PCM1792A you referred to earlier has only 66 levels (~6 bits) left at the D/A stage...

 

Edit: And actually if you used one of these parts for your tests, differences may be because of the converter architecture. At 42 dB attenuation the ICOB side falls of leaving only the SDM. Could be worth trying again with a pure SDM.

 

[Charles Hansen]

Allow me to express my appreciation that, in your early emulation of Linn, you did not adopt their product naming convention.

 

Thanks for the laugh!

 

When I was at Avalon, we gave our products names. This was such a difficult and time consuming process that I swore I would never do it again. So we use numbers at Ayre (actually letters plus numbers). The drawback there is that it is confusing sometimes. Even at the factory we are always calling the K-XR the "MX-R" and vice-versa.

 

There are some companies that have names that I think are really bad. But sometimes they get away with it. You just need to have something catchy. People used to say that the Beatles had a bad name, but it never seemed to hold them back!

 

[Charles Hansen]

Maybe the difference is here, since I don't use fixed point arithmetics, I never liked it. Everything is 64-bit floating point (11-bit exponent and 53-bit mantissa including sign)

 

I don't think that is the difference. If we are using 56 bit fixed point for our calculations, I don't think there is any meaningful difference between the two approaches.

 

After this, if there is still audible degradation with real attenuation using DACs, I know the problem is in the DAC electronics and should be fixed there.

 

What DAC are you using, and what DAC chip is inside?

 

And actually if you used one of these parts for your tests, differences may be because of the converter architecture.

 

Basically it sounds like you are saying that Burr-Brown's best DAC chip isn't good enough to hear your algorithms. So what DAC chips are good enough? Does a customer have to own a certain DAC to have proper results with your program?

 

[Miska]

What DAC are you using, and what DAC chip is inside?

 

The ones mostly in use; first has 2xAD1955, second one CS4398 and third one AK4396. Other things like CA DacMagic and various different PCI sound cards with different types of DACs are used less frequently.

 

Then I have couple of my old designs from early 90's in intermittent use. One has an early Crystal DAC, I don't remember the model right now and the second one is DF1700+PCM1700 and third one is DF1700+2xPCM63P.

 

Basically it sounds like you are saying that Burr-Brown's best DAC chip isn't good enough to hear your algorithms. So what DAC chips are good enough? Does a customer have to own a certain DAC to have proper results with your program?

 

IMO, Burr-Brown's best DAC chip is PCM1795, PCM1794A or PCM1704 depending on POV. But I recommend trying out also ES9018 or AK4399. And I have my strong reasons for choosing CS4398 for my latest design.

 

No, I was talking about your bad experiences with your algorithm.

 

I don't own any DAC where my algorithms would sound bad. I wouldn't release those if they would. I cannot say how they sound for others (equipment or ears), that's why anybody interested has to listen on their own. Then they can decide if they like my algorithms or not.

 

 

[jhbpa]

so maybe when Charlie says....?:

 

After my experiences with the design of our digital filter, I would never throw away resolution by using a digital volume control.

 

he could be a wee bit biased.

 

Carry On!

 

[barrows]

I would say it is Ayre's customers who are invested in preamps! Note that Ayre also makes a phono stage, and distributes a modified turntable in the US, so they really are obliged to make preamps (and they do produce very good preamps).

Of course Charles Hansen has his own opinion on this matter, but I would not suggest that he is expressing anything but his own, true, opinion-based on his listening experiences.

I own the mid level Ayre preamp, and my system sounds better using the preamp, than running direct and using Pure Music's volume control (only 10 dB of attenuation in Pure Music, with dithering). I always advise people that they need to audition in their own systems whether running DAC-Amp or DAC-Preamp-Amp is going to sound best-everything else is just theory and speculation. As Charles suggests, there is more going on in audio playback than we currently understand or can measure.

BTW, Ayre's K-XR Preamp offers a virtually no compromise adjustable gain stage to control volume (see John Atkinson's measurements at stereophile.com). An Ayre USB DAC incorporating the adjustable gain circuit of the K-XR (perhaps built to a more moderate standard) might offer an excellent alternative to digital volume control for those who do not need a preamp to accommodate multiple sources looking to run amp direct...

 

[Old Listener]

> The drawback there is that it is confusing sometimes. Even at the

> factory we are always calling the K-XR the "MX-R" and vice-versa.

 

Another permutation and you'll be saying K-Mixer. Now that would be a bad name for a DAC.

 

Hope that you are well and happy.

 

Bill

 

 

[jhbpa]

Thanks for the update on Ayre but be advised I don't need a lecture on the company's products. Since the 1990's I've owned:

 

V-3

V-1

K-3

K-1

D-1

C-7

C-5

A-7

QB-9

 

Most were great, a few (A-7 for example) were just good.

 

I still own a P-5, I've left out all the x's and e's for simplicity, but was actually the person who originally suggested the "x" designation to Gary Mulder (www.mulderaudio.com) who passed it on to Charlie.

 

In addition to the the two K's above, I've owned lots of high end preamps over the past 45 years, the only one that rivaled the K's being the BAT 52SE.

 

Point being that I always believed that removing the preamp had more minuses than plusses, possibly somewhat because Charlie said it was so and I respect him as one of the best two or three designers/developers in the history of high end audio.

 

That said, my belief system was fundamentally shaken when I acquired the Weiss DAC202. I'm preamp free forever. I'm also no longer auditioning DAC's (the QB-9 was a strong second). The DAC202 has also shaken all my preconceived notions regarding op-amps and discrete output stages.

 

I don't want to completely hijack the thread so I'll include my personal observation relevant to oversampling:

 

I don't understand very much of the technical discussion in this thread, but I my ears tell me that the software oversampling algorithms used by Pure Music dramatically improve the majority of my Red Book CD's. So my answer to the OP's question is Rob Robinson.

 

Barrows, as previously stated, you remain one of my favorite posters.

 

JB

 

[4est]

You really prefer oversampling with PM on your DAC 202? Can you tell me what you are upsampling to? You see, I have the red haired step child DAC2, and have not cared for the upsampling of redbook as of yet. I am more than willing to experiment some more though.

 

Are you 2x, 4x or straight to 192K? If I may ask, what is your computer configuration, as that may be my limitation as well.