Audio Theory Vs Implementation Reality

[jkeny]

I'm starting this thread as a branch from the "Conclusive "Proof"" thread to hopefully allow some interesting recent posts to be expanded.

That thread reported positive ABX results between 24/96KHz files & the downsampled version of those files (the details of this will be found in the original thread).
Responses to these reports hopefully give rise to an attempt at finding what is the underlying cause of the audible differences between these files.

Suggested explanations have been put on the table to account for this audible difference:
- ultrasonic energy in the Higher resolution files resulting in intermodulation distortion that can fold down into the audio range or be audibly noticeable
- Software that has been used to do the resampling has some flaws which introduce audible issues in the downsampled file

There may be other suggested explanations that I'm forgetting but so far I don't believe that any suggestion fully explains the positive results.

Underlying these suggested explanations is a premise that there is a disconnect between the mathematically perfect model definitions & their real-world implementations in software or silicon. This is particularly relevant & the commonest point of contention in digital audio.

I've seen it often suggested in response to people reporting hearing differences when a certain change/tweak is made - that it's because their system is broken in some way & that's why they are hearing the difference. It's a response that I find has more than a grain of truth to it while, at the same time, ignoring reality. My experience tells me that we are using systems (including software) that are not perfect & don't behave according to & produce results predicted by the mathematical models that underpin their operation. I presume that there are various approximations used for many reasons - expediency, economics or ignorance - that lead to results which can be measured & justified (within a certain framework).

So in the spirit of learning how things really are, I'm interested in hearing about the prevalence of such differences between the ideal models & their implementation. Also interested in the further, more difficult question, of the audibility of these difference as this is the final arbiter of their importance

Now some interesting recent posts on that thread touched on these points
- Tony Lauck did a mathematical/digital simulation test of a published 5th order sigma-delta modulator & found it's resolution to not come near 16bits. I woud be interested in hearing more about this & how it could be so wrong.
- Mention of the issue of noise modulation in sigma-delta DACs which I first came across in a presentation by the chief engineer of ESS at the time, Mallinson
- Considering that a DC sweep reveals such an issue it led me to consider why such a simple test result was not being reported for DACs & about datasheets, in general
- I recently came across measurements of the ES9018 flagship DAC from ESS showing this noise jump at about -35dB which seems to be different between the 8 output channels
- Furthermore it appears that digital clipping occurs when the oversampling filter is operational - a problem that was suggested is common to many S-D DACs

 

[Tony Lauck]

Digital Clipping in DACs can occur in upsampling DACs when playing files that have "inter-sample" peaks. The filter reconstructs the waveform and interpolates points that may be higher than the original digital signal. (Which we may assume are very close, but below full scale) So if one then takes the resulting signal and processes it without any gain reduction one may have distortion. The problem is not confined to oversampling DACs. It also occurs with non oversampling DACs that use filters, but any head room problems occur in the analog circuitry.

The problem is that gain reduction is only needed when playing some recordings, but if built into a DAC reduces the measured performance of the DAC and adds unnecessary noise to quiet recordings. Some DAC designers build in extra headroom to avoid digital clipping, but how much is enough? If one designs to the mathematical worst case this can be over 10 dB. If one goes for the worst case one will take a big hit on published specs and due to marketplace ignorance probably a big hit in sales. A better solution is to "punt" and push the decision on headroom off to the user. This is what Mytek has done with the Stereo192-DSD DAC that I have been using. It has a "volume trim" function that allows the user to have their choice of headroom. The extra headroom may be needed in a rock music studio, but is unneeded in a studio producing classical music.

A better solution would have been for the engineers at Phiips and Sony to have dealt with the problem of "inter-sample" peaks by providing a precise specification of how inter-sample peaks are calculated and what the acceptable limits are. To me, this lack of this specification demonstrates their carelessness or ignorance, or possibly lack of backbone in the face of management pressure. In the time frame the Red Book specification was being produced other mixed signal engineers were designing local area networks, such as the first commercial version of the Ethernet. The Ethernet specification came with detailed and precise specifications for the physical layer and what the digital to analog and analog to digital converters (the transmitter and receiver portions of the transceiver) had to do. The specification had a budget for filter ringing and took into effect things like the factor of 4 over pi that happen in peak signal amplitude when a square wave is turned into a sine wave after filtering out its harmonics. (I know this for a fact because I was on the team producing this specification and had many conversations with the designer of the physical layer.)

An even better solution would be to take the producers and mastering engineers responsible for the overly loud CDs out to the woodshed...

 

[jkeny]

Digital Clipping in DACs can occur in upsampling DACs when playing files that have "inter-sample" peaks. The filter reconstructs the waveform and interpolates points that may be higher than the original digital signal. (Which we may assume are very close, but below full scale) So if one then takes the resulting signal and processes it without any gain reduction one may have distortion. The problem is not confined to oversampling DACs. It also occurs with non oversampling DACs that use filters, but any head room problems occur in the analog circuitry.

Great information, Tony, as usual. Seems to be quite an oversight/error?

The problem is that gain reduction is only needed when playing some recordings, but if built into a DAC reduces the measured performance of the DAC and adds unnecessary noise to quiet recordings. Some DAC designers build in extra headroom to avoid digital clipping, but how much is enough? If one designs to the mathematical worst case this can be over 10 dB. If one goes for the worst case one will take a big hit on published specs and due to marketplace ignorance probably a big hit in sales. A better solution is to "punt" and push the decision on headroom off to the user. This is what Mytek has done with the Stereo192-DSD DAC that I have been using. It has a "volume trim" function that allows the user to have their choice of headroom. The extra headroom may be needed in a rock music studio, but is unneeded in a studio producing classical music.

The ES9018 has a register for master trim setting but this seems not to resolve the issue?

A better solution would have been for the engineers at Phiips and Sony to have dealt with the problem of "inter-sample" peaks by providing a precise specification of how inter-sample peaks are calculated and what the acceptable limits are. To me, this lack of this specification demonstrates their carelessness or ignorance, or possibly lack of backbone in the face of management pressure. In the time frame the Red Book specification was being produced other mixed signal engineers were designing local area networks, such as the first commercial version of the Ethernet. The Ethernet specification came with detailed and precise specifications for the physical layer and what the digital to analog and analog to digital converters (the transmitter and receiver portions of the transceiver) had to do. The specification had a budget for filter ringing and took into effect things like the factor of 4 over pi that happen in peak signal amplitude when a square wave is turned into a sine wave after filtering out its harmonics. (I know this for a fact because I was on the team producing this specification and had many conversations with the designer of the physical layer.)

An even better solution would be to take the producers and mastering engineers responsible for the overly loud CDs out to the woodshed...

Fantastic, in depth info. Thank you! Yes, it does appear that audio specifications didn't get the due considerations that signal engineers brought to other areas but as you say, products directed to the audio marketplace may well have different considerations, maybe more relaxed & less diligent specifications?

 

[microstrip]

(...) There are DAC architecture techniques that convert correlated noise into uncorrelated noise, at the cost of extra hardware and possibly a higher total noise floor. I suspect chip designers often make tradeoffs to make their product "measure" better at the expense of potentially sounding worse. (DAC designers of high end products marketed to subjective listeners don't commit this particular design "sin" since they eschew published measurements.) (...)

Could you give us a few examples of these architecture techniques? You are referring to a really interesting subject.

Post your question in jkeny's new thread and I'll be glad to answer.

Done.

 

[Tony Lauck]

Great information, Tony, as usual. Seems to be quite an oversight/error?

The ES9018 has a register for master trim setting but this seems not to resolve the issue?


Fantastic, in depth info. Thank you! Yes, it does appear that audio specifications didn't get the due considerations that signal engineers brought to other areas but as you say, products directed to the audio marketplace may well have different considerations, maybe more relaxed & less diligent specifications?

I would assume that trim register is what resolves the issue. The Mytek probably just passes the trim value configure in the user interface to the ESS chip. The Mytek uses a cheaper version of the ESS chip, but I suspect both versions come off the same fab line and the differences is just a price premium based on a noise-selection specification. (There may be newer versions that are different, as it has been some time since I investigated closely.)

The SABRE chip also has a 32 bit digital volume control. It can be used to achieve the same effect. Buy the way, I run my Mytek DAC direct to my amplifiers. I rely on the volume control in the DAC. When playing digital material I use the digital volume control built into the ESS chip, not the extra chip that provides variable gain op-amps. I have no use for unnecessary analog distortion. I have reasonable gain staging and seldom use more than 10 - 15 dB of digital volume control, except when listening to background music. The residual digital noise is far below the threshold of my hearing. If I had excessive gain to shed this method would not work as well and analog volume control might be better. In addition, there is no reason to work the op-amps in the output circuitry extra hard and make them distort more. I am not hearing any noise at my speakers when I play 24 bit digital silence, even if my ear is a few inches from the tweeter, regardless of the digital volume control setting.

One thing to keep in mind about that Mytek DAC. They may be marketing it to consumers on the "hi-fi" section of their web site, but the reality is that it is a pro-audio product and is best suited to more technically oriented people.

 

[LL21]

barely keeping up...well, actually not keeping up...but interesting reading in any event. Thanks for creating and keeping this thread going!

 

[Tony Lauck]

Could you give us a few examples of these architecture techniques? You are referring to a really interesting subject.

Here is what I originally wrote; "There are DAC architecture techniques that convert correlated noise into uncorrelated noise, at the cost of extra hardware and possibly a higher total noise floor." Two examples come to mind. Both involve the use of randomization so that a sub-system that is creating noise is not aware of the actual signal, and therefore the sub-system does not correlate its noise with the actual signal. At least, this is the idea behind these examples, the actual execution may not be perfect.

The first example is a technique that Barry Blesser described to linearize a DAC. (I believe it was in the 1980's.) It involves making multiple copies of the digital signal and adding random noise to each. The random noise is constructed so that it sums to zero. The multiple signals are sent to multiple DACs which separately create their analog output. The resulting analog signals are then summed to produce the final output. If the multiple DACs are perfect this will add no extra noise as the random noise will sum to zero in the final analog output. If there are any errors in the individual DACs, including otherwise benign gain errors, then there will be random noise in the output. In addition, the random signal added to each DAC may cause clipping unless the input signal is reduced in amplitude prior to adding the random noise. So even if the individual DACs are perfect and all the added noise cancels out there will be a net loss of resolution because of the reduced bit depth into the individual DACs. If good quality dither is used this noise will be uncorrelated (1st order: 0 or no distortion, 2nd order: 0 no noise modulation). The result will be less distortion or noise modulation than the original DAC had. (This worked with R2R DACs. I have no idea what would happen if delta-sigma DACs are used, as it would depend on the modulator algorithms.)

The second example is a technique that is used with low bit depth DACs that are constructed using thermometer code. (A thermometer DAC works like an old fashioned liquid thermometer, hence the name.) Here there are 2^n levels for an n bit DAC which are decoded using 2^n switches. The exponential number of switches limits the number of bits. So, for example, with some of the SABRE chips the number of switches may be 64 for a six bit DAC when one chip is run in 8 channel mode. This corresponds to a six bit DAC (running at a very high master clock frequency). When the DAC is configured to run in stereo mode then the number of switches per channel can increase to 256, giving an 8 bit DAC. (I could be wrong on the exact numbers, involved, but you get the idea.)

The reason for using thermometer code on a semi-conductor chip is that it is easy to make many transistor switches that are more or less identical, but hard to make devices that are precisely and deliberately different (as would be needed for an ladder DAC). A thermometer code DAC will be linear if all the switches are identical. (This is not exactly true if the individual switches are not perfect current sources, but this is a good approximation if they are powered from a stiff voltage source and the summing voltage is low because the summing resistor is small.) The problem is that the individual current sources may not be exactly identical. If the mapping of binary codes to activated switches is fixed by some hard wired logic then the resulting output will be non-linear, even though it may be monotonic.

The "dynamic element matching" technique consists of randomly permuting the assignment of switches to the thermometer code. This removes correlation between the input values that a particular switch processes. As a result the correlation between the input signal and the output signal is reduced. (The analysis is complicated.) The downside is creation of added noise. For example, suppose in a 6 bit DAC the input signal was a constant 32. Without any randomization 32 switches would be permanently on and 32 switches would be permanently off. The output would be a constant DC value. if dynamic element matching is added, then at any time 32 switches would remain on, but the particular switches involved would change, and this would add noise if the individual switches weren't perfectly identical. (There are other complications such as glitches created by differing timing between the switches and there are complex patented schemes to minimize these effects while keeping beneficial effects.)

Dynamic element matching is used in the ESS SABRE DACs and is described in more detail in the white paper. The white paper also provides a link to their patent which provides additional details of how it is done. http://www.esstech.com/PDF/sabrewp.pdf. I believe that dcS used similar techniques in their Ring DAC.

 

[jkeny]

Thanks Tony
Another way of improving linearity of the outputs - the continuous calibration technique used in some of the Phillips TDA DAC chips - TDA1549, TDA1543, TDA1545, TDA1387 where, as far a I understand it, the on-chip output gates are calibrated by charging up a capacitor through the current reference (all on-chip) & the charge on that cap then used for the output gate. The idea being that the fixed current reference charges each gate cap to the level that ensures the ref current passes through the gate - so differences between gate fets are dynamically matched in this way. It appears that switching feed through can be a issue.

I remember also Dr. Ulrich Brüggemann mentioning here another technique which attempted to address this correlated noise in another way. It involved converting the L-R signal prior to the D/A conversion into M-S & after D/A conversion back to L-R to share any data dependent distortions equally between both channels.

 

[Tony Lauck]

Another way to prevent distortion from being attributed to the foreground (instruments) rather than the background (ambient noise) is masking, specifically adding in broadband (white) noise. This may not be as crazy as it sounds if it is done in the right amount. It mimics an intrinsic part of the analog experience, where every playback of the same recording is different. This will not happen with a low noise DAC, rather all the noise on a particular recording will be reproduced on each playback. This may end up sounding unnatural. After all, if the musicians played the same song several times at the same microphones the random motion of the air molecules at the microphone diaphragms would be different each time.

While repetitious randomness won't matter to music lovers with a huge record library who seldom play a recording more than once, it may be important to audiophiles using a limited set of "reference" recordings to evaluate components and formats, etc... A long segment of white noise will sound random. A brief clip from this white noise that is looped will sound random at first, but after a few repetitions the listener will hear the pattern. (The noise must be white otherwise thumps due to DC offsets will be likely at the splice.)

These kinds of random variations in dither noise might be what Amir is using in his ABX testing....

 

[jkeny]

Was anybody at the AES conference papers session yesterday on Perception: Part 2?
I would be interested in hearing the details of the Meridian presentation "P14-3 The Audibility of Typical Digital Audio Filters in a High-Fidelity Playback System"
When are these papers published - there are a couple that look interesting although I doubt they will beat the information from Tony on this thread?

 

[wakibaki]

Hi jkeny.

Keeping dubious company, as ever? All you wishful thinkers trying to discredit conventional technology at the fringes huddling together for some mutual support? Don't you know all these theories are crackpot?

All that stuff about 'obstacles to discovery' is fine for pompous old inventors to spout at lunch long after they made their discoveries. And for crackpots to mutter to each other. You haven't invented squat yet.

You'll have to forgive me if I'm blunt, I'm no longer in remission and they say it's terminal.

You should go an get a proper engineering education instead of gazing wishfully through the restaurant window like a hungry little boy, if you could ever contain yourself from arguing with the lecturer. You can't work out your issues in a fistfight with the laws of physics. Particularly when you're just using it as a lazy way way out of accepting the discipline of truly understanding the subject.

Good luck.

w

 

[jkeny]

I'm sorry to hear that, Waki

 

[mojave]

A better solution would have been for the engineers at Phiips and Sony to have dealt with the problem of "inter-sample" peaks by providing a precise specification of how inter-sample peaks are calculated and what the acceptable limits are.

There is a standard, but it isn't limited to Redbook. ITU-R BS.1770 defines the algorithm used to determine if there are inter-sample peaks. This can be used both pre and post mastering. Playback software such as JRiver Media Center uses this algorithm in conjunction with its Volume Leveling to provide automatic true-peak correction to prevent any digital clipping. I have over 4,800 files in my library that need true-peak correction with the greatest correction needed on Eric Clapton's Reptile album that has peaks of +4.8 dB. I believe the standard was initiated by AES and later implemented into ITU BS.177.

 

[Tony Lauck]

There is a standard, but it isn't limited to Redbook. ITU-R BS.1770 defines the algorithm used to determine if there are inter-sample peaks. This can be used both pre and post mastering. Playback software such as JRiver Media Center uses this algorithm in conjunction with its Volume Leveling to provide automatic true-peak correction to prevent any digital clipping. I have over 4,800 files in my library that need true-peak correction with the greatest correction needed on Eric Clapton's Reptile album that has peaks of +4.8 dB. I believe the standard was initiated by AES and later implemented into ITU BS.177.

The ITU peak algorithm uses a 4x oversampling with a predefined low pass filter. This is not a true worst case for a DAC that uses a different oversampling rate or a different interpolation filter. This looks like a bunch of engineers trying to solve a problem that requires mathematics. If a DAC uses a FIR filter then the worst case intersample peak depends on the sum of the absolute value of the filter coefficients. The DAC designer can take appropriate steps to ensure adequate headroom. I use HQPlayer and it goes one better: it shows whether or not any digital clipping actually occurred during upsampling and if so the user can reduce digital gain prior to the filter. (This also works with my low frequency room correction as well.) Using all of the available digital headroom is, perhaps, relevant when working with 16 bit material, so there is some benefit from making these recordings as loud as possible (by gain, not dynamic compression). Given that all DACs today accept 24 bits of input there will be no loss of digital resolution by always doing an appropriate amount of gain reduction in the DAC itself when playing 16 bit material.

I can not comment, nor do I care, what happens with "loudness war" material. As far as I am concerned, a recording made this way has no musical value, because of demonstrated incompetence by the producers/musicians who approved the recording.

For reference, here is the ITU document. http://www.itu.int/dms_pubrec/itu-r/rec/bs/R-REC-BS.1770-3-201208-I!!PDF-E.pdf

 

[Atmasphere]

Buy the way, I run my Mytek DAC direct to my amplifiers. I rely on the volume control in the DAC. When playing digital material I use the digital volume control built into the ESS chip, not the extra chip that provides variable gain op-amps. I have no use for unnecessary analog distortion. I have reasonable gain staging and seldom use more than 10 - 15 dB of digital volume control, except when listening to background music. The residual digital noise is far below the threshold of my hearing. If I had excessive gain to shed this method would not work as well and analog volume control might be better. In addition, there is no reason to work the op-amps in the output circuitry extra hard and make them distort more. I am not hearing any noise at my speakers when I play 24 bit digital silence, even if my ear is a few inches from the tweeter, regardless of the digital volume control setting.

Just a FWIW: Most Op Amps of any merit will have distortion so low as to be inconsequential at most volume settings. Have you tried using it the other way 'round? Any difference?

 

[Tony Lauck]

Just a FWIW: Most Op Amps of any merit will have distortion so low as to be inconsequential at most volume settings. Have you tried using it the other way 'round? Any difference?

I did try the analog volume control of the Mytek. It did not sound as good as using the digital volume control. Perhaps if I had an external preamplifier I would have gotten different results. Some people like the analog volume control because of its slightly euphonic sound, but I want my DAC to be as transparent as possible as I use other means to "voice" my system.

Note that the Mytek has two volume controls, a "volume trim" function that is always in effect even with analog volume control and allows for different amounts of digital headroom, plus the volume control (which can be either analog or digital, user selected.) In addition I have a computer player that allows changing the level (32 bit digital) that gets sent to the Mytek. Best results come when all of these controls are used intelligently. When playing PCM files the best sound comes when I use HQPlayer to perform digital room correction, upsampling, and conversion to DSD128, sending that to the Mytek. The Mytek volume control works on DSD128 even when the volume control is set to be digital. (This function is done by the SABRE chip.)

 

[Atmasphere]

Another FWIW: A good volume control properly constructed will always outperform one that is digital.

Sensitivity to cables is another problem altogether (but can be the reason why the use of a good line stage after the DAC can actually be more transparent rather than less)! Many DACs (and preamps too) are quite sensitive so the cables can make a big difference. You would think that audiophiles would be interested in reducing the effects of the cables but IME this is not the case- its an uphill battle trying to convince them that its a good idea.

 

[LL21]

Another FWIW: A good volume control properly constructed will always outperform one that is digital...

What is your viewpoint on microprocessor controlled discrete resistor-ladder relays where there is something like a different relay for each volume level? I am no techie, so just interested to learn if you like the basic design concept. Obviously implementation/execution is key.

 

[Atmasphere]

We've tried it and spent a lot of money on it, I can say that. The quality of the relays is pretty important for such controls. Never got them to sound as good as a good quality rotary switch though.

 

[LL21]

We've tried it and spent a lot of money on it, I can say that. The quality of the relays is pretty important for such controls. Never got them to sound as good as a good quality rotary switch though.

Thank you...good to know. In my case, I will leave it to CJ who seem to be pretty experienced at what they do.