You know better than that, TNT... Correct me if I'm wrong! Your skepticism is foolish.
Edit: In my previous post I meant 0.05% resistors should sound no different from 0.01%. 0.05% THD maybe someone can hear but I did a test on 0.1% and couldn't tell the difference. Granted the test was only one short piano track I believe, so the THD might show better in other materials. Now, if you can prove, offer convincing evidence, or even just a coherent theory on why 0.05% resistor should sound worse than 0.01% (I have the 0.01/0.02% version myself), I would take back everything I said.
Btw what does "losing in" even mean in English? It sounds really weird... like some combination of succumbing to one's obsessions and going against all the facts
Have you tried poetry?How does white noise help measure ripple?... I thought the old ripple calculations are all worst case estimates, i.e. ripple when switching from all resistors off to all on.
Last edited:
Nvm zfe... Your measurements were taken from the Vref directly! Wow
If we were give in to our obsessions and start optimizing everything, your measurements showed that there is no reason to go beyond 6800uF assuming there is no mains noise interfering. Quite a strong conclusion... Looks like 2200uF polymer caps for the rev.4/5 would be perfect. Not that we know if it's needed at all...
WowIf we were give in to our obsessions and start optimizing everything, your measurements showed that there is no reason to go beyond 6800uF assuming there is no mains noise interfering. Quite a strong conclusion... Looks like 2200uF polymer caps for the rev.4/5 would be perfect. Not that we know if it's needed at all...
Last edited:
If I were to take a wild guess and assume that the Vref mod from stock rev.4 ~350uF to something like ~5000uF adds to the originally anemic low end, before we have any ABX evidence, how could this possibly be explained by the lower ripple? In white noise playback it seems more Vref capacitance translates to the ripple levels in lower frequencies flattening to around the levels of other frequencies, which is quite flat in the audible range.
Could it be that the linearity of the nonmonotonic bits really do contribute to the sound directly?
Edit: factory mod (rev.2) has about 4uV (-128dbfs relative to -7.5dbfs 4v) ripple when playing white noise (at 0db?). 6800uF reduces this by about 12db (4x) in everything except the low end. 3uV difference is about the voltage in the 20th bit. Not sure if this is convincing evidence... On the other hand, could the relatively higher low end ripple (still lower than without mod) contribute to a more pronounced bass? No easy speculation.
Could it be that the linearity of the nonmonotonic bits really do contribute to the sound directly?
Edit: factory mod (rev.2) has about 4uV (-128dbfs relative to -7.5dbfs 4v) ripple when playing white noise (at 0db?). 6800uF reduces this by about 12db (4x) in everything except the low end. 3uV difference is about the voltage in the 20th bit. Not sure if this is convincing evidence... On the other hand, could the relatively higher low end ripple (still lower than without mod) contribute to a more pronounced bass? No easy speculation.
Last edited:
Maybe it was mentioned, but what is your listening setup to evaluate the dac?
A repost from zfe: the red traces are the ripples in dbfs. The first is 0.1R vs 1R with no/minimal capacitance (factory config in rev2+) load resistor, the second is 0.1R vs 1R with 6800uF.
I don't know why Paul ended up saying that his low res mod (0.01R?) doesn't sound good, but it seems 0.1R is at least much much better than factory 1R in all aspects.
The results align perfectly with Paul's simulations. The simulations predicted that with large capacitance, e.g. 6800uF, replacing 1R with 0.01R would reduce the low end ripple by a very large amount and bring it much closer to the rest of the spectrum. He even tried 0.008R which yielded even better results. I would guess that 0.1R in this case still isn't as good as 0.01R with large capacitance, though maybe more stable with no capacitance?
But what matters is if we can actually hear the difference and if so what it translates into. It's hard to even speculate if we don't have that data. TNT, if you still think that Soren isn't staying out of this because he believes none of it matters in listening, maybe you should share your insights with us.
P.S. I thought the shift registers contribute 0.3R to the Vref? Where did the internal resistance go? I suppose what we see here is what the resistor ladders get and goes straight to output.
I don't know why Paul ended up saying that his low res mod (0.01R?) doesn't sound good, but it seems 0.1R is at least much much better than factory 1R in all aspects.
The results align perfectly with Paul's simulations. The simulations predicted that with large capacitance, e.g. 6800uF, replacing 1R with 0.01R would reduce the low end ripple by a very large amount and bring it much closer to the rest of the spectrum. He even tried 0.008R which yielded even better results. I would guess that 0.1R in this case still isn't as good as 0.01R with large capacitance, though maybe more stable with no capacitance?
But what matters is if we can actually hear the difference and if so what it translates into. It's hard to even speculate if we don't have that data. TNT, if you still think that Soren isn't staying out of this because he believes none of it matters in listening, maybe you should share your insights with us.
P.S. I thought the shift registers contribute 0.3R to the Vref? Where did the internal resistance go? I suppose what we see here is what the resistor ladders get and goes straight to output.
Attachments
Last edited:
Zfe, would you be willing to procure some 0.01R 0603 resistors and repeat the measurements for it with large capacitance? If you read Paul’s blog, which I’m almost certain you have, his simulations suggest 0.01R should be best absent 0.008R which may be impossible to get...
Also, have you heard any differences? Did you have a chance to record the outputs? I think lots of us (minus our volatile TNT, I won’t ever even try speak for him again..) would be interested in contributing to ABX tests and report subjective findings, if any. The tough part is having someone who can mod relatively easily and have good recording equipment....
Again, I think this is still uncharted territory since we’ve had zero ABX result so far. The methodology isn’t perfect but I think if we are to try anything, it’s probably worth a shot. Soekris is great and pretty much perfect from the audio engineering perspective given what we currently know, except for the dual-mono issue sadly. I’m very happy with a dual-mono 0.012% dam1021 and 5000uF per vref rail. Can certainly go without any more improvements, but I find our little explorations pretty interesting.
We’re not correcting for a error, but expanding our understanding to go further.
Also, have you heard any differences? Did you have a chance to record the outputs? I think lots of us (minus our volatile TNT, I won’t ever even try speak for him again..) would be interested in contributing to ABX tests and report subjective findings, if any. The tough part is having someone who can mod relatively easily and have good recording equipment....
Again, I think this is still uncharted territory since we’ve had zero ABX result so far. The methodology isn’t perfect but I think if we are to try anything, it’s probably worth a shot. Soekris is great and pretty much perfect from the audio engineering perspective given what we currently know, except for the dual-mono issue sadly. I’m very happy with a dual-mono 0.012% dam1021 and 5000uF per vref rail. Can certainly go without any more improvements, but I find our little explorations pretty interesting.
We’re not correcting for a error, but expanding our understanding to go further.
Last edited:
First the picture of the modded part.
View attachment 496509
As most of the vias are covered by the "green paint" at my DAM, I had not the connet-form-below-option, I soldered the capacitors (16V 2700uF Panasonic FM) directly to the main buffer cap. And as my case is not high enough, I had to lay them flat.
The green caps are 360pF WIMA FKP2 soldered directly to the GND and Vcc pins of the shift registers. This was for the 250MHz oscillations (weather they were now in the DAM or oscillations in my oscilloscope), surely they not harm (but the physical connection is a bit fragile ... possibly I will end to remove them again in the future).
So to the sound. I have to admit, that I have not listened to my other modifications a) as I had not always enough parts lying around to modify all rails. b) I was very tight in time and only did the full modification for the most promising version. So I compare only with the unmodified DAM.
I currently only can listen with my headphone (an "vintage" Stax Lambda Pro + SRM1 Mk2). First observation is, with high sample rate recordings, the unmodded DAM and this headphone I sooner or later got some kind of headache. With the modded version this seems to be gone! I suspect that either that 39kHz noise hump either caused distortions it the head phone amp or this high frequency junk in the signal wereas transmitted to well to the ear and although inaudible caused some pain. So good point
The sound has definitely changed very noticeably!!
I hear more details, the sound is even more than before only "there" (in the sense that you don't think it comes from the headphones).
An prominent example is the bass section where in some piece I suddenly hear a very deep drum in the background with clear modulations of its vibrations (before this was more some unarticulated noise). Moreover, with enough volume up, I felt the bass part in the body as never with the headphone before... more like with speakers.
Impulses of piano or guitar snares have more attack and are more well defined impulses, again not covering details.
This changed the sound of this setup that much, that I have to listen a lot more to see if I get "used" to that. I can not say that the tonal balance got wrong. Definitely there are "improvements". But I am not yet decided If I just have not heard the "truth" before or if the effects are overemphasizing at some point and there is now to much of ... many things.
I encourage the people here with more analytic or professional ears to give this mod a try. I would be glad to hear other opinions.
Especially in comparison to the other mods.
Apologies for missing this post a few pages after your measurement, zfe. I think I've read this post before so I may be subconsciously influenced but my observations match yours exactly, even though my mods are on stock rev.4. Also seems like you don't have a ton of time testing different mods either, or have fond memories of 0603 soldering... Still I think Paul's later recommendations on 0.01R might be worth looking into, if there aren't other complications that I overlooked.
Hmm, 1 word clock cycle is 1.3uS in x8... 1 bit clock is 32x smaller... Soren it would be great if you can be more specific on your prediction of how the clocks lock
I was lying in bed and couldn’t stop the train of thoughts. For one thing, maybe TNT has had the right methodology all along - listening, measuring and recording after each change over a period of days. Unfortunately, (and I say so without pleasure) it seems he has not made any progress on the DAC itself but focused on other parts of his system. TNT also has a kind of reluctance to theorize. Perhaps it’s unfair to attribute it to the weakness of his mind, which has proven to be more sound than mine on many occasions over the past couple of years. That aside, why hasn’t he tried though?
I was wondering if the notion of our ears being more sensitive does indeed have some truth in it. I do not for a moment believe that our auditory system has more resolution than our instruments, the demonstration of which I dare anyone to try to tell apart a sine wave with 0 THD and 0.001% THD - easy for cheaply available soundcards, hard for us mortals. Soren said on other forums that he believes the advantage of his DAC is in time domain, not frequency. Is it just an excuse for relatively unspectacular measurements? If not, why is it so hard to come up with good tests?
TNT has also been stressing the imprortance of component sound coloration, which is total nonsense if you look through our measurements. Ceramic caps will always outmeasure TNT’s ears (though maybe not his imagination at times, I’m sure he can get a good laugh out of this too), but it’s really not inconceivable that they do distort the sound in a way that is significant yet hard to detect.
So, my brilliant world-changing idea is this: what if we’re missing analytic tools, not instruments? What if our ears seem to pick out differences not because of our superior auditory cells (which are still cool but probably no match for semiconductors in some cases), but because of the processing that follows? I’m not talking about our most vivid audiophilic imagination, but of our ability to pick out sounds and analyze texture, etc. Even the best musicians don’t always hear in terms of the frequencies. Though some might be able to identify tone frequencies with some accuracy, it’s still a rare and relatively rudimentary skill. Now, I should’ve done a bit more digging before I suggest anything but I assume no one here will judge me negatively for it that much considering what everyone has to deal with on forums... Is there a chance that we can make use of the breakthroughs in neural networks in recent years, including RNNs that is essentially a time-domain model, and the more recent transformers that are so impressive that they basically replaced everything overnight? I haven’t had any experience with audio but this seems a semi-reasonable proposal. Does this agree with people’s intuitions?
I was wondering if the notion of our ears being more sensitive does indeed have some truth in it. I do not for a moment believe that our auditory system has more resolution than our instruments, the demonstration of which I dare anyone to try to tell apart a sine wave with 0 THD and 0.001% THD - easy for cheaply available soundcards, hard for us mortals. Soren said on other forums that he believes the advantage of his DAC is in time domain, not frequency. Is it just an excuse for relatively unspectacular measurements? If not, why is it so hard to come up with good tests?
TNT has also been stressing the imprortance of component sound coloration, which is total nonsense if you look through our measurements. Ceramic caps will always outmeasure TNT’s ears (though maybe not his imagination at times, I’m sure he can get a good laugh out of this too
), but it’s really not inconceivable that they do distort the sound in a way that is significant yet hard to detect. So, my brilliant world-changing idea is this: what if we’re missing analytic tools, not instruments? What if our ears seem to pick out differences not because of our superior auditory cells (which are still cool but probably no match for semiconductors in some cases), but because of the processing that follows? I’m not talking about our most vivid audiophilic imagination, but of our ability to pick out sounds and analyze texture, etc. Even the best musicians don’t always hear in terms of the frequencies. Though some might be able to identify tone frequencies with some accuracy, it’s still a rare and relatively rudimentary skill. Now, I should’ve done a bit more digging before I suggest anything but I assume no one here will judge me negatively for it that much considering what everyone has to deal with on forums... Is there a chance that we can make use of the breakthroughs in neural networks in recent years, including RNNs that is essentially a time-domain model, and the more recent transformers that are so impressive that they basically replaced everything overnight? I haven’t had any experience with audio but this seems a semi-reasonable proposal. Does this agree with people’s intuitions?
Did a quick search and it turns out people actually tried to use FIR in ANNs for time-domain analysis but that’s with very ancient multilayer perceptron. There’re also some voip audio quality analysis using ANN though the architecture is not so exciting and they’re testing 13kbps against 64 so not the ideal audiophile resolution.
Anyways, something like this could be useful for audio engineers but I doubt anything will be published on this topic anytime soon, just seems too trivial and frivolous... So I won’t hold my breath for anything but still it’s something really interesting to think about in spare time!
Correction: people in the CS community have written about memes generation and the like, so maybe nothing is too frivolous for us.
P.S. Music transformer is not a new name for fancy audiophile transformers: Music Transformer: Generating Music with Long-Term Structure
Anyways, something like this could be useful for audio engineers but I doubt anything will be published on this topic anytime soon, just seems too trivial and frivolous... So I won’t hold my breath for anything but still it’s something really interesting to think about in spare time!
Correction: people in the CS community have written about memes generation and the like, so maybe nothing is too frivolous for us.
P.S. Music transformer is not a new name for fancy audiophile transformers: Music Transformer: Generating Music with Long-Term Structure
Last edited:
P.P.S. On a second thought, even if something like this were to work, a non-transparent system (i.e. black-box, not what you have in mind for cheap speaker setups...) still doesn’t have the predictive power that theory offers... sad but still an interesting/educational idea.
That said, maybe professional audiophiles can be replaced once and for all!
That said, maybe professional audiophiles can be replaced once and for all!
Last edited:
The measurements (e.g. the -128dBFS) are in the frequency domain. Ripple is something in the time domain.
To get the RMS value of the ripple you need to sum the bins of the frequency domain in the range of interest. The RMS value displayed in the lower left corner is the sum over all bins (RMS of the "ripple" over full measurement bandwidth).
For the pp-ripple you need to do the inverse Fourier transform and look at Min/Max, or do the measurement again and display it in time domain (scope-mode).
TI in one of their publications on ADC was only interested in reducing ripple insofar as THD is concerned. Maybe it’s different for DAC, or maybe TI was wrong. In any case, everything has both a time and frequency domain. If we assume that there is something of interest in the time domain (or frequency domain) that we can’t measure, there are only a few possible experiments on the Soekris platform I think.
1. Replace 1R with 0.01R. Paui’s simulations showed woolly waveforms in 500hz with 1R, but that goes away with 0.01R, amplitude of 500hz ripple also goes down dramatically, not that it may matter in whitenoise RMS. The data gathered could be Vref frequency domain measurements to compare against existing. And maybe time domain if oscilloscopes have that accuracy. In addition, output recordings would help resolve to some extent whether such a change, if confirmed by measurements, in indeed audible.
2. Measure (and listen to) 1121. The shift registers seem different and possibly better even if it’s still the same LVC type.
3. Correct for resistor error. There’s still a time component to the glitches.
4. Prototype a board with the “buffers” Soren mentioned at one point between shift registers and the ladder to significantly lower the impact of internal resistance of the shift registers. This might lower the speed according to Soren, but not sure how much that would impact performance. This would be quite complex. Even if we offer to help fund the research effort, I’m not sure if Soren is enough of the adventurous type to find interest. In comparison, option 3 is much easier, almost no prototyping cost except for some software dev, and promises at least better looking frequency domain measurements, even if those figures don’t really mean anything.
PM me if something takes off. I’d be happy to contribute financially. It’s unlikely that we’ll get an NSF grant for this
1. Replace 1R with 0.01R. Paui’s simulations showed woolly waveforms in 500hz with 1R, but that goes away with 0.01R, amplitude of 500hz ripple also goes down dramatically, not that it may matter in whitenoise RMS. The data gathered could be Vref frequency domain measurements to compare against existing. And maybe time domain if oscilloscopes have that accuracy. In addition, output recordings would help resolve to some extent whether such a change, if confirmed by measurements, in indeed audible.
2. Measure (and listen to) 1121. The shift registers seem different and possibly better even if it’s still the same LVC type.
3. Correct for resistor error. There’s still a time component to the glitches.
4. Prototype a board with the “buffers” Soren mentioned at one point between shift registers and the ladder to significantly lower the impact of internal resistance of the shift registers. This might lower the speed according to Soren, but not sure how much that would impact performance. This would be quite complex. Even if we offer to help fund the research effort, I’m not sure if Soren is enough of the adventurous type to find interest. In comparison, option 3 is much easier, almost no prototyping cost except for some software dev, and promises at least better looking frequency domain measurements, even if those figures don’t really mean anything.
PM me if something takes off. I’d be happy to contribute financially. It’s unlikely that we’ll get an NSF grant for this
I got curious again and was finally able to make out the difference between B and C files on this jitter page: jitter_1. The archived version shows that C has 300ps 300hz jitter compared to the reference B: jitter_2results. I do worry that the site might have hidden agendas and as a result may not be trustworthy. But it's a good test.
The difference is pretty minute. The jitter file at 2.0s with perhaps the F# note is noticeably muddier/smoother than the reference file. The reference file has more texture while the jitter-injected file just kind of smooths the edges of the pluck. (Guess MSB was full of s*** when they say jitter=harshness)
I guess this shows to some extent that maybe 1021 isn't as good as it can get. But I don't believe that the differences can be heard without good ABX setup. On the more interesting side, as I was going through the test over and over, my hearing was significantly habituated to the sound to the extent that I can no longer tell the difference. A short break brought back the sensitivities in my hearing. This phenomenon itself is well-documented in scientific literature, but what's interesting is that after habituation, the sounds seem to have lost their texture, leaving behind only the pitch which matched perfectly between the two files. It reminded me that I was taught in middle school that the three properties of sound are pitch, timbre and loudness. If timbre can be separated from pitch, what can THD, IMD, J-test, etc. tell us about proper timbre reproduction? We don't listen to square waves either, so I can't see that much value in square wave impulse measurements. Is this obvious to everyone else or am I just wrong?
The difference is pretty minute. The jitter file at 2.0s with perhaps the F# note is noticeably muddier/smoother than the reference file. The reference file has more texture while the jitter-injected file just kind of smooths the edges of the pluck. (Guess MSB was full of s*** when they say jitter=harshness)
I guess this shows to some extent that maybe 1021 isn't as good as it can get. But I don't believe that the differences can be heard without good ABX setup. On the more interesting side, as I was going through the test over and over, my hearing was significantly habituated to the sound to the extent that I can no longer tell the difference. A short break brought back the sensitivities in my hearing. This phenomenon itself is well-documented in scientific literature, but what's interesting is that after habituation, the sounds seem to have lost their texture, leaving behind only the pitch which matched perfectly between the two files. It reminded me that I was taught in middle school that the three properties of sound are pitch, timbre and loudness. If timbre can be separated from pitch, what can THD, IMD, J-test, etc. tell us about proper timbre reproduction? We don't listen to square waves either, so I can't see that much value in square wave impulse measurements. Is this obvious to everyone else or am I just wrong?
Attachments
(PDF) Study paper for Timbre identification in sound
http://www.cs.northwestern.edu/~par...nd Timbre in the Auditory System - shamma.pdf
https://www.ee.columbia.edu/~dpwe/pubs/MuEKR11-spmus.pdf
Particularly the sections on Timbre.
booksdescr.org
Skim chapter 1 and 5.
Extra credit assignment that came to mind:
No Really: Teach Your Toddler Perfect Pitch | Psychology Today
If you have a toddler in the house, train him to identify 2nd, 3rd, and 4th harmonic distortions - the 1khz tone is of the greatest importance, use it most lavishly in the training. It only works when he's young, so the onus is on you to prepare him for a profiting career in audiophilia when he grows up! Ability to identify those -110db harmonics will surely come in handy when you want to sell those ESS dacs
http://www.cs.northwestern.edu/~par...nd Timbre in the Auditory System - shamma.pdf
https://www.ee.columbia.edu/~dpwe/pubs/MuEKR11-spmus.pdf
Particularly the sections on Timbre.
booksdescr.org
Skim chapter 1 and 5.
Extra credit assignment that came to mind:
No Really: Teach Your Toddler Perfect Pitch | Psychology Today
If you have a toddler in the house, train him to identify 2nd, 3rd, and 4th harmonic distortions - the 1khz tone is of the greatest importance, use it most lavishly in the training. It only works when he's young, so the onus is on you to prepare him for a profiting career in audiophilia when he grows up! Ability to identify those -110db harmonics will surely come in handy when you want to sell those ESS dacs
Last edited:
Chapter 8 is even more relevant and worth a more careful read.
Also, a quick search of TotalDAC internals reveals HC594 and HC574, the latter being a d-type flip-flop buffer for the former shift register, a method that Soren brought up in the discussion earlier. I'm now inclined to believe that TotalDAC does outclass Soekris. Soren believed wrong when he made his design decision - Vref ripple does matter, even if it's for reasons we don't fully understand (the book chapter suggests that there might be as many as 34 dimensions to timbre, whereas there's only 1 each for pitch, i.e. THD, and amplitude, i.e. frequency response. I feel an urge to map our wealth of audiophile knowledge to the proposed dimensions, but it seems analogous to using phrenology to guide cognitive neuroscience... "That silver gold oil capacitor sounds richer and has sharper attack!"). Btw, it also seems that TotalDAC uses 1 giant electrolytic cap for each Vref rail? diyAudio R&D is catching up!
----------
Soren, is there a chance that we can crowdfund you to build a simple lookup table into FPGA for non-linearity correction, perhaps optionally added to one or more of the filter banks? Can you give us a project cost estimate?
Also, a quick search of TotalDAC internals reveals HC594 and HC574, the latter being a d-type flip-flop buffer for the former shift register, a method that Soren brought up in the discussion earlier. I'm now inclined to believe that TotalDAC does outclass Soekris. Soren believed wrong when he made his design decision - Vref ripple does matter, even if it's for reasons we don't fully understand (the book chapter suggests that there might be as many as 34 dimensions to timbre, whereas there's only 1 each for pitch, i.e. THD, and amplitude, i.e. frequency response. I feel an urge to map our wealth of audiophile knowledge to the proposed dimensions, but it seems analogous to using phrenology to guide cognitive neuroscience... "That silver gold oil capacitor sounds richer and has sharper attack!"). Btw, it also seems that TotalDAC uses 1 giant electrolytic cap for each Vref rail?
diyAudio R&D is catching up!----------
Soren, is there a chance that we can crowdfund you to build a simple lookup table into FPGA for non-linearity correction, perhaps optionally added to one or more of the filter banks? Can you give us a project cost estimate?
Last edited:
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.