Hi Mark, am interested in this board, I can't PM as I am in moderation at the moment, could you PM me on how I can get one from you.
Many Thanks
Wow, a lot of information! I agree with Thorsten on so many things. Ferrites, different types of noise, etc.
Regarding noise, there are some interesting comments by Bart Kosko on different types of white noise and what they look like in the time domain:
https://www.edge.org/response-detail/11715
So, what does anyone think it means that there are different kinds of white noise, all of which have a flat frequency distribution?
Seems to me they must differ in their phase distributions.
IME they can sound different too. Apples and oranges.
Regarding noise, there are some interesting comments by Bart Kosko on different types of white noise and what they look like in the time domain:
https://www.edge.org/response-detail/11715
So, what does anyone think it means that there are different kinds of white noise, all of which have a flat frequency distribution?
Seems to me they must differ in their phase distributions.
IME they can sound different too. Apples and oranges.
Last edited:
Getting back to all of the problems with noise, FPGAs, ground bounce, etc., only dacs I know of that are designed to minimize those sorts of problems are those of Andrea Mori. I think Marcel has figured out what Andrea is doing, but I don't think I should describe the architecture in detail here. That said there is a pic of an earlier version than I am using at TheWellAudio website:
Image taken from https://www.thewellaudio.com/twsdac-dsd/
Image taken from https://www.thewellaudio.com/twsdac-dsd/
I designed a PCB to interface a JLSounds V3 to Marcel's Valve DAC, I'm sure I have some spares here, here being only 50miles or so from you.
I included an isolator for the DSD-on flags that Marcel's DAC required - the isolated flags on the JLSounds V3 are low for DSD-on but Marcels Valve DAC required it to be high so I used the flag from the non-isolated side of the JLSounds. IIRC the RTZ DAC is the same.
I also included a facility to show the DSD sample rate, using four bi-colour LEDs - red for 44.1KHz family rates and yellow for 48KHz.
It all worked perfectly for the Valve DAC project - the module is in the bottom right of the first image below.
Last edited:
Have you thought about devices such as that below. I have used these in other applications. These are analog gates that can isolate the supplies from the input digital and also can isolate the ground if the analog output is connected to dual supplies. In other words there is too much that can go on as unpredictable on the digital side (shoot through being one). With such isolators the current feeding the resistance networks are sourced from the analog side through its own supply.
https://assets.nexperia.com/documents/data-sheet/74HC_HCT4051.pdf
A gaussian distribution as reducing the amplitude of the tails is what I am thinking about in terms of an I/V. It is considered that digital signals of the form FIRDAC can be viewed from an RFI/EMI perspective that generates the DC shifting and thereupon IM, as seemingly the dominant mechanism in folding frequencies into the audio band.
Because of the issues you and others had with DSDON and MUTE polarities, I made the polarities switchable by mounting the right 0 ohm resistors in the latest version of the RTZ shift register DAC.
I designed a PCB to interface a JLSounds V3 to Marcel's Valve DAC, I'm sure I have some spares here, here being only 50miles or so from you.
I included an isolator for the DSD-on flags that Marcel's DAC required - the isolated flags on the JLSounds V3 are low for DSD-on but Marcels Valve DAC required it to be high so I used the flag from the non-isolated side of the JLSounds. IIRC the RTZ DAC is the same.
I also included a facility to show the DSD sample rate, using four bi-colour LEDs - red for 44.1KHz family rates and yellow for 48KHz.
It all worked perfectly for the Valve DAC project - the module is in the bottom right of the first image below.
PM sent
I think there are less exotic ways to end up at a non-Gaussian distribution. Just take Gaussian noise and pass it through something non-linear and you have non-Gaussian noise. Besides, the central limit theorem doesn't apply when you add only a small number of equally-distributed components with finite variance, or when you multiply rather than add them (the result then tends to a log-normal rather than a normal distribution).
I think you are probably right about the phases, at least for the case of a not too small discrete Fourier transform. The inverse transform boils down to adding a bunch of sine waves. If those sine waves had equal amplitudes and phases that are uniformly distributed over 2 π rad, the sum would rather quickly converge to Gaussian noise, according to Bennett's 1948 article. With a different distribution of the phases, you can get something very different. For example, in the continuous case, a flat magnitude response and a phase of 0 (all cosines) results in a Dirac impulse at t = 0.
Regarding audibility of noise with different probability density functions, see https://www.diyaudio.com/community/threads/high-order-dither-listening-test.313257/#post-5207912 , with apologies for the methodological error.
Last edited:
I made a simulation of Marcel's Firdac as close to the real world as possible.
The only thing I did not take into account is the eventual Vref ripple on the shiftregisters caused by current-peaks while changing the flip-flop's output state, But apart from disruptions caused by ultra short current peaks, the way signals and shiftregisters are interleaved may prevent changing Vref from playing a significant role.
I used a repetitive 69h or 01101001 input signal which is generated in a subcircuit and used a 10Mhz BCK, almost equal to DSD256 and LTSpice operated at a 0.1nsec resolution.
Small subcircuit boxes were inserted at vital places to give a 5 nsec delay on tr and tf, since the digital components in LTSpice have no delay.
The shiftregisters are subcircuits made from D flipflops, each having a same delay box added at their outputs.
The 3k01 Firdac resistors are already made part of this subcircuit.
In the attachment below signals taken at various positions are level lifted in the image to get a better view on each individual signal.
From top to bottom:
Input signal: continuous 69h
V(outn): output from Firdac 1
V(outp): output from the Firdac 2
V(CM): Common Mode signal of both output signals showing a perfectly stable signal
V(Diff): Differential Mode signal whith twice the amplitude as the SE signals.
From supposedly CM products is no proof whatsoever at this stage.
So implementing cross-coupling or adding a second set of Firdac's seems a waste of time
Hans
The only thing I did not take into account is the eventual Vref ripple on the shiftregisters caused by current-peaks while changing the flip-flop's output state, But apart from disruptions caused by ultra short current peaks, the way signals and shiftregisters are interleaved may prevent changing Vref from playing a significant role.
I used a repetitive 69h or 01101001 input signal which is generated in a subcircuit and used a 10Mhz BCK, almost equal to DSD256 and LTSpice operated at a 0.1nsec resolution.
Small subcircuit boxes were inserted at vital places to give a 5 nsec delay on tr and tf, since the digital components in LTSpice have no delay.
The shiftregisters are subcircuits made from D flipflops, each having a same delay box added at their outputs.
The 3k01 Firdac resistors are already made part of this subcircuit.
In the attachment below signals taken at various positions are level lifted in the image to get a better view on each individual signal.
From top to bottom:
Input signal: continuous 69h
V(outn): output from Firdac 1
V(outp): output from the Firdac 2
V(CM): Common Mode signal of both output signals showing a perfectly stable signal
V(Diff): Differential Mode signal whith twice the amplitude as the SE signals.
From supposedly CM products is no proof whatsoever at this stage.
So implementing cross-coupling or adding a second set of Firdac's seems a waste of time
Hans
Attachments
Yes, but we aren't always applying statistics to only a simplistic case of noise (although in this case we were). How about for different noise mechanisms arising from different physical effects? Say, resistor resistor excess noise from substrate surface defects versus from end-cap attachment method? Or, how about for audibility thresholds for untrained versus trained listeners on a better system that was historically used for threshold studies? We might find thick tails are empirically a better fit than gaussian, say, before we study further to find out if nonlinearity is the most useful process model to explain why?
My only point here is that I think we should keep an open mind about the more general applicability of the findings described by Kosko in the linked article.
I not so sure the result was as much about human audibility thresholds of dither noise as much as it was about typical dac reproduction quality, as well as other common system problems such as "grainy" and or "veiled" sound due to ground loops and or other ground noise problems. IOW and IMHO its probably the stuff that is harder to quantify, such as for example, noise skirts, which may account for audible differences between the best dacs and best reproduction systems, as opposed to consumer equipment that measure well in typical AP tests.
Moreover, I think it is pretty well accepted that some people can hear the difference between CD dither algorithms without cheating by turning the volume way up during very low level passages. Its not a uncommon ability among some recording and mastering engineers. It was also one of those mastering engineers who demonstrated to Bruno Putzeys that class-D output inductor hysteresis noise/distortion could be audible to some people on some systems.
Last edited:
That seems like an overkill when playing 16 bit CD’s having a max S/N ratio of ca 96dB which seems perfectly adequate when properly dithered.
That we can achieve higher figures just like one digit PPM THD figures is impressive but not really needed IMO.
Hans
P.S. Just as a remark, SNR relates to power where S/N relates to voltages.
If they are properly dithered then the effective audible bit-depth for human listeners can extend below the top of the noise floor. Could be more like 18-bits of audible music information. Why add even more noise and distortion to what is already there? In addition, we aren't always playing CDs. There is some hi-res in my collection that I would like to hear without the compression often needed with CD audio to help hide dither and or truncation effects.
The compression and limited headroom of CDs is easily seen by comparing CDs and Hi-res versions of the same original recording. Modern LUFS volume metering will show average and peak dynamic range, as well post-reconstruction intersample overs. IME hi-res versions typically show more peak headroom below 0dBFS and more dynamic range than CD.
Last edited:
Replace RMS measurements with peak. Have a look. Dither especially,
Of course Tape or LP using peak....
The main issue is that digital noise is not really random, at best it is pseudo-random but often has distinct pattern designed to average out with a sufficient long window.
But the human hearing is not a THD analyser, FFT or massive averaging system.
So what is MEASURABLE and what is AUDIBLE are interestingly different.
And then there is the observation that listening is a learned skill. Those who have gone to see the elephant commonly react very differently to certain sounds than those who did not.
For example, the time domain discrimination for direction of percussive sound for human hearing appears to suggest > 100kHz equivalent bandwidth yes few people ever can hear a steady state 20kHz sine wave.
While with a string quartet a 10kHz lowpass in not very audible, try that with a Gamelan orchestra from Bali (I have an Ex GF from Bali).
Thor
For what it's worth, my valve DAC with its -85 dB(A) to -93 dB(A) noise floor sounds fine to me and to everyone else who heard it. I wouldn't use it with digital volume control, though.
Regarding noise, the noise of high-dynamic-range sigma-delta DACs is often largely analogue. Reducing analogue noise usually costs much more power and chip area than reducing digital noise, so most of the error budget goes to analogue.
Regarding noise, the noise of high-dynamic-range sigma-delta DACs is often largely analogue. Reducing analogue noise usually costs much more power and chip area than reducing digital noise, so most of the error budget goes to analogue.
I changed the resistors of second stage (OPA1678) to have same gain as your original filter. Distortion at 10kHz improved only marginally (about 0.3dB). Sorry, no more measurements from me on this thread so you will just have to take my word on this
EDIT: Since level was now 6dB lower the marginal improvement may be due to just having lower distortion in ADC.
That's interesting, as it refutes my second-stage-distortion hypothesis.
Unsurprising. Many people get used to listening to satellite radio with its lossy compression and think it sounds fine. Maybe its that they hear what they expect to hear. IMHO that's not justification for the SQ of satellite radio. In reality it sounds awful to someone who hears what the lossy compression is really doing.
Of course, I'm not saying that your dac sounds like satellite radio. I'm just trying to point out that sounding fine is relative to whatever people are used to.
Last edited:
Hmmm, No. But are we sure that an unloaded FIR DAC output with various parasitic in the logic IC's, which may or may not be modelled correctly.
How are you modelling the inputs and outputs? Ideal logic, simple static model or dynamic models with dynamic impedance's?
What are you doing to account for the whole analogue stage? Nothing?
Anyway, I already said you guys of "nothing here" win. No need to spend more time.
Honestly, I am not sure what you are simulating and why. But it is nothing like the Sim's I am running and I would not expect the same results. And you don't need to convince me, I already admitted defeat.
What question is your simulation attempting to answer and how can you be sure it will give the right answer? Example, we could look at the differential RTZ signal top make the common mode issue attested in the literature on the topic in fact exists in our sim and then look at the DAC and make sure it removed this.
Just saying, not really interested TBH.
Enjoy the DAC you have. I do that with TDA1541 and will get a PCM56 DAC for downstairs.
These already sounded PERFECT (for me) three decades ago, "when I wore a younger man's clothes..."
Never thought I'd fit into a Billy Joel song... Or I'd be living next to a Beach not far from "The Beach".
"But at my back I always hear
Time’s wing'ed chariot hurrying near;
And yonder all before us lie
Deserts of vast eternity."
Thor