Designing ES982x-based ADC - feedback wanted

I understand your points. Really all I am aiming for here is something where the Pi captures the source to digital format, in a way that when I play it back the average person will say "it works". It would be nice to also be able to show that THD+N figure is better than existing solutions. This is not aimed for use in audiophile systems like the IanCanada stuff. My equipment that will feed to it is just regular consumer stuff anyway.
 
Some more quotes from Dr. Geddes:
...Our conclusion; people are satisfied with THD and IMD. It’s like the story of the cop who asks a drunk under a street light what he is doing on his hands and knee’s. The drunk replies “I’m looking for my car keys.” The officer asks “Where did you loose them?” and the drunk replies “Over there by my car.” Baffled, the officer asks “Then why are you looking for them here?” to which the drunk replies, “Because the light is better.” Everyone knows that THD is meaningless, but it’s easy to do and “the light is better.”

...The bottom line here is that we know so little about how humans perceive the sound quality of an audio system... that one should question almost everything that we think we know about measuring it. From what we have found most of what is being done in this regard is naive. Things like distortion measurements that don’t consider masking... Maybe someday in the future we will be able to quantify perceived sound quality and move audio away from a marketing dominated situation to a data driven one.

http://www.gedlee.com/Papers/Comments on howard.pdf
 
... Maybe someday in the future we will be able to quantify perceived sound quality and move audio away from a marketing dominated situation to a data driven one.

http://www.gedlee.com/Papers/Comments on howard.pdf

Perhaps, but OTOH most people buy a sound system by listening to it.

Btw I like that analogy of the drunk looking for keys. I instinctively agree, not from listening experience (I don’t have enough) but because math says that there are a lot of distortions that will be totally invisible in THD+N graph.
 
Last edited:
Perhaps, but OTOH most people buy a sound system by listening to it.
If the alternative is THD+N, SINAD, etc., then couldn't blame them. Unfortunately the hi-fi stores of yore where people could go to listen have been replaced by big-box discounters and Amazon. No more listening, unless perhaps you have a friend who will let you listen at the friend's house. For those near Auburn, CA, USA, I have offered let people visit and listen whatever we have going at the moment if they would like.
 
If the alternative is THD+N, SINAD, etc., then couldn't blame them. Unfortunately the hi-fi stores of yore where people could go to listen have been replaced by big-box discounters and Amazon. No more listening, unless perhaps you have a friend who will let you listen at the friend's house. For those near Auburn, CA, USA, I have offered let people visit and listen whatever we have going at the moment if they would like.
Here in Australia stores with listening spaces are definitely still around if you go looking for them.
 
You might want to read "Designing Audio Power Amplifiers, 2nd Edition," Chapter 16, "Other Sources of Distortions." Fourteen distortion mechanisms are described in detail.

Beyond what Cordell has identified in the above book, Geddes has shown that linear distortion can be audible in some circumstances, there are a couple of AES papers about audible distortion mechanisms in line and headphone cables due to EM field forces, ESS has described training its executive team to notice 'state variable settling' during quiet periods immediately following a loud transient, and Purifi as described an audible speaker distortion mechanism that listening tests in collaboration with a university showed naïve listeners did not notice.

Of course there must be more.

Regarding the above, please don't claim that anything not shown to be audible in a published study can safely be assumed to be inaudible. We all should know better by now.

Regarding THD+N, most commonly these days there is an acquired dataset of samples which may or may not contain evidence of certain distortions and or noises depending on how and when the dataset was acquired. Another issue is the common practice of discarding phase information before analysis which could then result in some audible effects not showing up or else looking smaller than they really are. Windowing can further complicate some types of analysis.

By the time everything is lumped into one number, THD+N, almost all meaning has been lost. Yes a lot is lumped into that number, but a hash algorithm can similarly lump a lot of information into one number. Of course trying to make meaningful sense out a a hash of an acquired sample dataset seems silly. THD+N turns out to be not very useful either.
 
Last edited:
Regarding the above, please don't claim that anything not shown to be audible in a published study can safely be assumed to be inaudible. We all should know better by now.
Why disregard published studies? Of course they may be exceptional individuals but no one has shown up yet in controlled tests.

By the time everything is lumped into one number, THD+N, almost all meaning has been lost. Yes a lot is lumped into that number, but a hash algorithm can similarly lump a lot of information into one number. Of course trying to make meaningful sense out a a hash of an acquired sample dataset seems silly. THD+N turns out to be not very useful either.
Why do you keep pushing this silly notion of "one number, THD+N"? Do you have any links to somebody using just a single number for evaluation? Besides I already said there are loads of other measurements as well.
 
We are talking a lot about one number right now, THD+N. Seems like one member who sells commercial ADCs tends to quote that number a lot for his product. Just sayin'

OTOH, completely agree there are lots of other measurements. IIRC Cordell does not describe how to measure every one listed in his distortion chapter. ESS didn't describe how to measure state variable settling either. So seems doubtful we have good ways at our practical disposal to measure some things.

Regarding published studies, I'm not disregarding them offhand. I quoted what Geddes published about what we know about audio system measurements. The problem isn't with published studies so much as it has been with how engineers have misinterpreted such studies as giving exact hard limits, and or giving complete, highly-reliable, highly-linear models of human hearing. How else could they calculate over at ASR exactly what level of dac distortion must be inaudible to all humans?
 
Last edited:
Why disregard published studies? Of course they may be exceptional individuals but no one has shown up yet in controlled tests.


Why do you keep pushing this silly notion of "one number, THD+N"? Do you have any links to somebody using just a single number for evaluation? Besides I already said there are loads of other measurements as well.

The quote about the drunk searching for his keys was about using just THD+N. And marketing materials for consumers (and even some high-end stuff) often have that as the headline measurement (and sometimes the only one), which is what I think @Markw4 is lamenting. I guess it's like sports cars - you can't measure how it feels through a corner so everybody just quotes their top speed.
 
We are talking a lot about one number right now, THD+N. Seems like one member who sells commercial ADCs tends to quote that number a lot for his product. Just sayin'
Just sayin' still waiting for your alternative metric. If you think Geddes metric is it put your money where your mouth is and provide a Geddes metric your dac. Don't be shy. Be a trailblazer!
 
Looked at it a bit yesterday. It requires measuring the transfer characteristic of the DUT (which can be done if phase information is retained), then deriving a twice-differentiable function to model the transfer curve. Appears frequency dependence (a 'memory' effect) can be taken into account if some feedback is applied the model's static nonlinearity blocks. Other types of memory behavior and or more than 'small' nonlinearity may pose difficulties/complications. There is a thread on the subject and some back and forth discussion. A spreadsheet was posted to help with the calculations. Can provide links if anyone is interested. Will continue to think about it..
 
Whatever you do with any kind of distortion measurements etc. you're never going to be able to do more than what an ADC / DAC is capable of doing. Also, power supplies can be an SMPS / linear regulator followed by a filter and shunt regulator if necessary.

If you pick an SMPS / linear regulator, pick one whose control bandwidth exceeds 20kHz. Further, if you pick an LDO to cleanup the SMPS output, pick one with a control bandwidth greater than the SMPS's switching frequency. Use shunt regulator with proper filtering and bypassing for voltage references.

THD+N is a basically useless term when it comes to data conversion, as the distortion largely depends on the amplitude of the signal. Unfortunately, full scale signal is not what we see all the time. All you can do is to provide the system with the electrically cleanest / quietest environment practically possible.
 
THD+N is a basically useless term when it comes to data conversion, as the distortion largely depends on the amplitude of the signal. Unfortunately, full scale signal is not what we see all the time. All you can do is to provide the system with the electrically cleanest / quietest environment practically possible.
It seems that you are missing the point. All ADC/DAC manufacturers specify their chip performance using THD+N. That makes it the easiest metric to verify that implementation is doing justice to the chip. Other measurements (spectra, jitter, phase noise ...) and listening tests should be used to further verify that implementation is successful. E.g. if manufacturer specifies THD+N to be -120dB at 0dBFS from 20-20kHz and your device reaches barely -100dB you can tell that the implementation is not very successful assuming that datasheet level performance was the goal.
 
Agreed, maybe you could say (looking at THD) if something is not right, but that would be the only use for that metric since datasheet values are often measurements obtained using a particular system.

For example, many DACs from Burr Brown use AP's System Two that offers several filters of which, (if I remember correctly), a steep 20kHz filter is used for THD+N measurements. Other manufacturers have similar setups to sweeten their datasheet values.