I don't recall referring to an 'it' - I believe I said that S-D DACs were noisy. Admittedly I could have written 'S-D DACs are noisy modulationy' but seems too cumbersome in English.
How did you go about measuring this -115dBC level you mention? If you used some averaging then bear in mind the noise tends to be impulsive, its crest factor will be rather high. ISTM our ears are sensitive to its peak level, not average.
Are you able to duplicate the FFT shown on my blog - a pic I gleaned from Head-Fi? Incidentally how do you estimate that an average level of -115dBC (i.e. 10dB of noise modulation) is 'pretty innocuous' ?
I believe the flaw in the claim that multibit S-D DACs can be 'perfect' is in the way he goes about measuring them - with FFTs. He presents an FFT where the noise floor is unvarying with signal level and then claims that shows an absence of noise modulation. However the FFT shows only the average noise so its not sufficient evidence for the absence of noise modulation - the noise can still be changing while its average level remains constant.
However I mentioned that paper for their work on DSD.
I would echo this. Just because a method of measuring is convenient, does not mean that then sufficient is known about behaviour. A semi-random, aberrant behaviour may be almost invisible in the frequency domain, but be clear as day in a snapshot of the time domain history - and you can bet Murphy's Law that the ears will pick it up ...
Perhaps "have too much modulation noise" would have been better. Just saying noise suggests it easy to see. Clearly it isn't or it would have been addressed most likely.
The -115 dB is referenced to the peak levels in the individual bands of the excitation signal. Not the peak level or full scale. FWIW I'm using 13100 point FFT at 48 KHz. I will try various methods of capturing what is coming out to look for intermittent peak noise levels.
I have been trying with no success to reproduce what they show. Perhaps you can provide a link? Maybe there is a detail I'm missing. Feeding 1 KHz at -35 dBFS and then -35 dBFS (and most of the levels around -35) how no noise floor changes beyond those that happen because its noise. I'm using single captures (no averaging) and peak holds in the bins with averaging with no success.
Rather than eyeball estimations of noisy baselines I'll post some pictures that will help. My point was actually that signals consistently below -115 dB are not that important. No transducer I know of could possibly reproduce them. And, checking my calibrations this evening, the actual level was -130 dBFS. I will be more careful in the calibrations going forward. The 4 tones are all selected to have no common harmonic or IM relationships so they will be maximum stress in that sense.
Most of the current DAC chips use multi-bit modulators. Except for ESS and AKM I think we won't see new DAC or ADC chips. Its too much money to develop for too little return. The business is small low power chips that can get -100 dB THD+N for less than $2 and codecs with local DSP.
Picture 1- -35 dBFS and -36 dBFS superimposed. Picture 2 low frequency multi-tone at two significantly different levels, -20 dBFS and -44 dBFS. Both pictures; one capture 131,072 points, 48 KHz, no averaging. Source Picture 1 Altor digital generator, picture 2 Praxis via EMU1616M. Capture in both cases EMU 1616M and Praxis.
Attachments
Analog vs PCM vs DSD
In the same room, between the speakers might not be too bad, away from room modes which concentrate in corners, out of the dispersion pattern of most speakers, a not bad compromise, given lack of sophisticated vibration isolators.
I see all this dumping on vinyl… Yet it lives on not purely through inaccuracy, I believe, but like all analog systems, quantum resolution underneath all the noise and distortion mechanisms. Being the highly complex nonlinear systems we are, we are affected by that to some degree (I know this is unproven).
My recent casual test: I had terrible wow/flutter with my first new turntable in a long time. Tried to measure it, and I ran into off-center wow which dominates everything (but we learn to filter it out). Replaced cartridge and it went away. It was arm cartridge resonance. I made some digital recordings (Alesis Masterlink 24/96) with the bad cartridge, and the digital recordings sounded more or less ok. I made those recordings with the music playing.
Hence far out idea: the time resolution of 96kHz digital just isn't sufficient. Either we need more total information, or we need information in between certain dots that isn't there when a 96kHz grid is laid down.
Other idea: to determine audibility, don't use nice music, use something with a far out problem.
For a brief moment I was thinking DSD with the hypothetical 2.8mhz resolution would be way to go. Listening to DSD demonstration convinced me of the reverse. Nothing sounded good. I think it messes with the time information too much in the name of reducing noise (and it has too much noise in the first place). Slow grid better than that, or maybe just more pleasant.
DSD now seems to have something like noise gating effect that all pwm-like systems seem to have, even without actual noise gates. Noise modulation, as has been stated here. John Atkinson plotted "jitter" noise modulation tests of Sony XA777ES for CD jitter test and pure tone for SACD. The CD jitter test shows spikes but the SACD reproduction of 10kHz brought an audible spectrum wide 10dB higher background noise (-118dB). If you look at the standard -90dB signal noise tests, at 10kHz the CD and SACD reproduction have identical noise. So there's obviously a noise modulation process going on, and it modulates spectrum wide noise.
Sony SCD-XA777ES multichannel SACD/CD player Measurements part 2 | Stereophile.com
Here's my compilation of criticism regarding DSD:
Audio Investigations: Hirez PCM is better than DSD
Now this has made be very suspicious of s-d dacs also, though that may be less warranted. In JA's test above, the CD layer was of course being reproduced by the same S-D dac. But what would an R2R DAC do?
In the same room, between the speakers might not be too bad, away from room modes which concentrate in corners, out of the dispersion pattern of most speakers, a not bad compromise, given lack of sophisticated vibration isolators.
I see all this dumping on vinyl… Yet it lives on not purely through inaccuracy, I believe, but like all analog systems, quantum resolution underneath all the noise and distortion mechanisms. Being the highly complex nonlinear systems we are, we are affected by that to some degree (I know this is unproven).
My recent casual test: I had terrible wow/flutter with my first new turntable in a long time. Tried to measure it, and I ran into off-center wow which dominates everything (but we learn to filter it out). Replaced cartridge and it went away. It was arm cartridge resonance. I made some digital recordings (Alesis Masterlink 24/96) with the bad cartridge, and the digital recordings sounded more or less ok. I made those recordings with the music playing.
Hence far out idea: the time resolution of 96kHz digital just isn't sufficient. Either we need more total information, or we need information in between certain dots that isn't there when a 96kHz grid is laid down.
Other idea: to determine audibility, don't use nice music, use something with a far out problem.
For a brief moment I was thinking DSD with the hypothetical 2.8mhz resolution would be way to go. Listening to DSD demonstration convinced me of the reverse. Nothing sounded good. I think it messes with the time information too much in the name of reducing noise (and it has too much noise in the first place). Slow grid better than that, or maybe just more pleasant.
DSD now seems to have something like noise gating effect that all pwm-like systems seem to have, even without actual noise gates. Noise modulation, as has been stated here. John Atkinson plotted "jitter" noise modulation tests of Sony XA777ES for CD jitter test and pure tone for SACD. The CD jitter test shows spikes but the SACD reproduction of 10kHz brought an audible spectrum wide 10dB higher background noise (-118dB). If you look at the standard -90dB signal noise tests, at 10kHz the CD and SACD reproduction have identical noise. So there's obviously a noise modulation process going on, and it modulates spectrum wide noise.
Sony SCD-XA777ES multichannel SACD/CD player Measurements part 2 | Stereophile.com
Here's my compilation of criticism regarding DSD:
Audio Investigations: Hirez PCM is better than DSD
Now this has made be very suspicious of s-d dacs also, though that may be less warranted. In JA's test above, the CD layer was of course being reproduced by the same S-D dac. But what would an R2R DAC do?
On the surface this violates information theory, a signal bandlimited at 20kHz sampled at 96kHz has no missing information. Vinyl has so many mechanical artifacts I doubt any judgement about digital quality will come from digitizing LP's.
No, a processing delay of exactly one sample period. If the signal is bandlimited to 20kHz it's value is known exactly (within the limits of noise and jitter in the A/D) between samples.
More information violates Nyquist.
BTW I have little or no interest in the 1 bit DAC controversey. I have used and purchased 24/96 external devices like field recorders and external sound cards and I have listened to digital played through SY's 24/96 DSP. There is nothing more in quality I would make an effort to achieve in the digital only domain.
Last edited:
So if I sample a 20Khz sine, then do it again with no sync to the 96K, the two digital streams should be exactly the same?
jn
I am grateful for a learned critique, which is hard for me to get except in places like this.
Where is the 20kHz band limit of which you speak? Even if the LP were band limited, the arm-cartridge resonance is producing higher sidebands than were recorded. That's the idea of finding a situation where something isn't right to test for audibility.
Now the other place you might apply a hypothetical 20kHz bandwidth is in my hearing, and indeed I cannot hear pure tones much above 16kHz. But I believe that is not the whole story. There is stuff going on in my highly nonlinear and complex processing that is affected by higher frequencies in some way. As if, say, I have sampling at 32kHz but it's not in purely linear time. It has a different pattern, that makes it sensitive to even higher frequency information that doesn't exactly fit that pattern.
Thinking about this, I took a look at Shannon on Wikipedia. But Shannon works in terms of probability functions…and it seems to me they could be just about anything. WHat I may be questioning is not Shannon per se but particular derivatives like Shannon-Hartley. And I wonder if they are making all the right assumptions for complex nonlinear systems like me. And that's where I am now.
I think the answer is that the reconstructed analog will be exactly the same. If you have a few minutes this is a really good primer video on digital audio: Xiph.Org Video Presentations: Digital Show & Tell I still find 24/176 and 24/192 sounds better than the 16 bit versions but it just may be limitations in the processing (or my firmly imbedded expectation bias).
To sample twice you need to reconstruct the analogue signal before the second sampling. That requires a reconstruction filter to throw away images, which will also do double-duty as the anti-aliasing filter for the second sampling. The second digital stream will be fully equivalent to the first digital stream, in that it encodes the same original analogue information. It is most unlikely to be identical to the first digital stream, unless it has a synchronised clock and the intervening filter has a delay exactly equal to N times the sampling interval.jneutron said:
I'm sure you know this, so why ask?
Now that's where the time resolution comes in. The "nonlinearity" in the time is several orders of magnitude below even the most optimistic numbers for temporal/spatial resolution. It's even more orders of magnitude below time uncertainty in mechanical systems (e.g., LP, tape).
I did not mean reconstruct then re-sample. My question was geared towards a single 20 Khz sine being fed into two independent 96k converters which are not synched to each other.
jn
As an example.
Assume an analog signal being fed to a left and a right converter.
The converters are 1/4 hz off.
The two datastreams are then reconstructed using two synched 96Khz D/A's.
The two waveforms will time dither at the 5 uSec level, at 1/4 hz rate.
jn
Yes, the reconstructed signals are exactly the same (again within noise, etc. margins). The phase difference could just as well be in the 20kHz signals and the same 96k sampler.
EDIT - Now you're talking a totally different story, .25Hz off in respect to what? Sampling at 96000.25 and reconstructing at 96000. and you don't have 20kHz anymore. Sample rate conversion in general loses bit-perfectness. I also don't know of any system where the A/D's are not synced so this seems a bit contrived.
Last edited:
In a multitrack studio with a 64 channel mixing board, is every signal processing function analog? If not, are all the channels A/D's with local clocks, or is there a master?
I'm actually not worried about 1/4 hz at 20Khz, nor at 1Khz. I would worry about maintaining ITD though.
jn
- Status
- Not open for further replies.