advertisement


Noise Shaping

Jim Audiomisc said:
Sony's SBM was, but I don't know about the DG system.
Click to expand...
There are others. I understand the BMG label uses Apogee's UV22 system (see here). There's also EMI's ART (Abbey Road Technology) mastering, which I suspect is a name for their own high-order noise shaping.
​
Jim Audiomisc said:
FWIW one of the postings here lists a NS set of coefficients that may be better than my simplistic 1st order,

https://hydrogenaud.io/index.php/topic,47589.0.html

but as yet I've not had a chance to try it.
Click to expand...
There are useful minimum audibility noise shaping coefficients of several orders in the annex of "Minimally Audible Noise Shaping" by Stanley P. Lipshitz, John Vanderkooy, and Robert A. Wannamaker, J. Audio Eng. Soc., Vol. 39, No. 11, November 1991 for anyone who wants to try them. I did try some time ago to wade through the mathematics in a Gerzon-Craven noise shaping paper to compute my own noise shaping coefficients, but it was simpler in the end to use some pre-computed ones in the software I was writing at the time.
 
Ciunas Audio said:
Just going back to the original topic - noise shaping - is there a possible overlooked issue with this - a fluctuating noise floor?

I remember ESS chief designer Martin Mallinson giving a presentation in which he spoke about the DS modulators producing non-periodic steady state noise artifacts - see here

Apparently, this DS modulator behaviour can be seen when noise Vs DC offset is plotted
Click to expand...
This is a very entertaining presentation. It's a long time since I practiced digital signal processing professionally but the key issue discussed in it seems to be about the performance of one-bit delta-sigma modulators used in DACs, not specifically about noise shaping. One-bit delta-sigma modulators are known to have problems and to be imperfectible in principle. There are AES papers about this such as "Why 1-Bit Sigma-Delta Conversion is Unsuitable for High-Quality Applications", from 2001. ESS seem to have found a way to minimize the one-bit delta-sigma problem, but using multi-bit delta-sigma converters seems to be a more fundamental solution.

Ciunas Audio said:
Is this modulator separate to the noise-shaping function or part of it?
Click to expand...
I think the delta-sigma modulator's problems are a separate issue. Although the two functions are often connected since noise shaping can be applied to the error feedback loop of a delta-sigma modulator. I am not aware of any particular problem in noise shaping per se such as signal-correlated noise floor (but do note that I have been out of the field for some time). ​
 
davidsrsb said:
Multi bit DS does not solve the problems, just reduces them
Click to expand...
Thanks for the correction. I read Lipschitz and Vanderkooy's:
"In contrast, multi-bit sigma-delta converters, which output linear PCM code, are in principle infinitely perfectible. (Here, multi-bit refers to at least two bits in the converter.) They can be properly dithered so as to guarantee the absence of all distortion, limit cycles, and noise modulation. The audio industry is misguided if it adopts 1-bit sigma-delta conversion as the basis for any high-quality processing, archiving, or distribution format to replace multi-bit, linear PCM."​

in a different way. Perhaps it's a technical language problem since it's many years since I was in the research lab. I am now not certain I have interpreted the AES paper ("Why 1-Bit Sigma-Delta Conversion is Unsuitable for High-Quality Applications") properly. I will read it again.
 
The papers I linked earlier in this thread on DS show that multibit is also capable of artifacts. If this was not the case the Wolfson/Cirrus and BB/TI multibits would measure and sound better than ESS
 
The complication lays in the number of 'states' the system can get into and how to analyse, etc, them.

Simply 1-bit systems have AIUI the inevitable problem of not being completely dithered, so can be expected to show a problem like idlers or lockup. SACD sets a max modulation well below the full possible range to try and avoid this, but then depends on partial dithering. The snag being that Philips and Sony kept secret the details of the modulators they used. But when I analysed ones they'd published I could find systematic problems. I did exchange code and info with the designers, and we agreed this happened, but they they said they used other designs which I was not allowed to check.

Higher order 1-bit systems can exhibit 'orbits' under given conditions, and you can't always predict them. So have to find them by sufficiently exhaustive tests on all their internal states *and* input sequences. Which means some may reside unfound...

FWIW This relates to a topic I got into for quite different reasons myself some decades ago - the concept of 'quasi-chaotic' behaviour as either a noise source or for encryption/steganography. All being well, I'll write about some of that elsewhere in due time, but can't say all of it. Not audio, anyway!

Multibit systems tend to be fixable simply because you can apply sufficient dither. Only snag then is ensuring the dither is suitably 'noise like'... which goes back to the conundrum of what makes a deterministic sequence 'noise'. 8-]

BTW I did look at the Lipshitz set of coefficients, but IIRC they were designed for something different to 96k/24 -> 96k/16 so weren't directly ideal for that. Didn't experiment with how to shift the band where the process noise would be minimised.
 
davidsrsb said:
The papers I linked earlier in this thread on DS show that multibit is also capable of artifacts. If this was not the case the Wolfson/Cirrus and BB/TI multibits would measure and sound better than ESS
Click to expand...
Yep, this in particular "Noise modulation is well understood for PCM, yet there is no well established theory for even low-order sigma–delta modulators"which appears to signify that it is more art than engineering & I wondered if we are facing the same lack of rigour in noise shaping techniques - Jim, seemed to be suggesting this early in the thread?

Jim Audiomisc said:
"Multibit systems tend to be fixable simply because you can apply sufficient dither. Only snag then is ensuring the dither is suitably 'noise like'... which goes back to the conundrum of what makes a deterministic sequence 'noise'. 8-]"
Click to expand...
I either don't understand this as relates to noise modulation issues in SD modulators or I disagree with it?
 
There is a general problem of "where did the dither sequence come from?" in processes that are entirely digital, because they a 'finite state' systems. i.e. You can only have a finite set of possible values in all the locations in the system. And this includes the 'source' of the 'noise'. Think of something like a max-length sequence generator. 'max' can't be infinite, just (expletive) looooong. And the 'noise' is actually a deterministic sequence. You can 'fold' the input into that, and that shuffles things a bit, but the same problem still lurks.

The details then depend on the size/shape of the system. (Number of bits per value, number of locations, etc.) But mean that without some 'real noise' you may encounter problems. So in reality the aim is to make them as rare and minor as possible.

In audio we're helped because more good recordings will contain some genuine background noise, and that in itself helps to 'dither' the processess near the start of a chain. So the universe can, for free, provide us with a decent dithering noise source. *Provided* the capture can resolve that above the 'noise' level it, itself requires as a minimum.

I should perhaps confess that noise sources and measurement were something I got into for a while, and 'noise' turns out to be quite a slippery concept. However in my case the noise and 'fake noise' tended to be at 100 GHz not audio. You may recall when 'Chaos' was popular. At the time I was interested in what I used to call 'semi-chaotic' behaviour. i.e. looks like noise, but has a hidden pattern you can use. This work was eerily parallel to examining systems like DSD and some forms of digital processing. But the good news is that people can and do design well-arranged ADCs, etc, that generally avoid the problems. The trade-off is that low-order and low-bit systems are easier to analyse, but more prone to problems, whilst high order systems are harder to analyse but less prone if higher-bit. There is no absolute
guarantee card for that with a 'new' system, though. :)

I'll see if I can dig out some of the results I got from the Sony 5-th order DSD modulator design they published, but then say they didn't use.
 
But what happens when the "noise" is actually 31.778 kHz computer monitor line scan or just above audio smps contamination? You can detect these on some sample 88k and 96k recordings.
 
Thanks, Jim, for your detailed reply.
I believe the problem I have with your post about dither & the reply above is that you seem to be considering this noise purely in digital "'finite state' systems", viewpoint. I may be wrong & would welcome clarification.

I also welcome your simulation results of the Sony DSD modulator design & it's consequent problems.

The noise I was referring to is not confined to this purely digital viewpoint - it is the analogue noise (possibly variations in ground currents?) which are loosely correlated to & derived from the processing of the digital signals & becomes part of the analogue output in a D/A system. Some example of this would be ground bounce withing ICs, Inter symbol interference, settling time of voltage outputs, Mallinson's "non-periodic steady state noise artifacts", etc.

In some sense, I'm saying the same as davidsrb but, whereas he's talking about electrical noise that is injected into the digital system, I'm talking about electrical noise that is generated internally from within the digital processing itself.
 
Jim Audiomisc said:
Multibit systems tend to be fixable simply because you can apply sufficient dither. Only snag then is ensuring the dither is suitably 'noise like'... which goes back to the conundrum of what makes a deterministic sequence 'noise'. 8-]
Click to expand...
That (your first point about being able to apply sufficient dither in multi-bit systems) is how I read the Lipscihtz and Vanderkooy paper. One of the papers cited by davidsrb (I read both yesterday) does seem to contradict this saying that the fixes are only partial. It may be that I haven't yet understood the specific contexts around the different statements.

And yes - the requirement is that the dither is sufficiently noise-like is certainly there. Now that I think about it that might be one of the points I should look at in the paper that seems to question my original assumption.

Not having studied the issue of "noise-like" I was also wondering about the differences in being noise-like between a good but deterministic sequence generator and one that modifies the same basic generator with entropy gathered from the computer's operation (with or without entropy pool re-use). I think the first is deterministic for a given seed and the second is not deterministic. How much does the non-deterministic nature matter WRT being noise-like, I wonder?

I know someone of whom to ask this from a cryptographic PoV but not a signal processing PoV.

Jim Audiomisc said:
BTW I did look at the Lipshitz set of coefficients, but IIRC they were designed for something different to 96k/24 -> 96k/16 so weren't directly ideal for that. Didn't experiment with how to shift the band where the process noise would be minimised.
Click to expand...
Yes - I failed to say that I was using the filters for changing resolution at 44.1 ksamples/s - not the same as your situation. ​
 
Generating very long pseudo random sequences is easy and takes little silicon. The snag is that synchronising to the stream becomes hard. Not a problem if you just want to make "noise" but no good if you are trying to scramble data to be recovered
 
davidsrsb said:
Generating very long pseudo random sequences is easy and takes little silicon. The snag is that synchronising to the stream becomes hard. Not a problem if you just want to make "noise" but no good if you are trying to scramble data to be recovered
Click to expand...

Also, a long sequence gives more space for a section to go 1010101010101.... for a fair time.

8-]
 
OK, these may be of interest

http://jcgl.orpheusweb.co.uk/temp/DSD1.pdf

http://jcgl.orpheusweb.co.uk/temp/DSD2.pdf

They are two *partial rough drafts* of something I never finished. Worked on them back in 2004/5. The significant part here is the last graph in DSD2. You can see the 'noise level' over a wide band waggles about in an odd way. Some kind of cyclic behaviour. The problem is that with high order low bit systems it is difficult to predict when these may occur, so matter of luck if you find them, or even notice them. Never got further. The Philips people ran my code, said they couldn't see anything wrong with the code, and got the same behaviour. But then said they used different modulators anyway... whose details they would not disclose. That made them a little difficult for me to model. :)
 
BTW There was a Penguin Guide to Chaos and an Electronics World article as well as a New Scientist one on some of the 'chaos and noise' things I took and interest in. You may be able to find info on the web, but I've not yet got to the relevant time period in my 'biography' pages. Too busy arguing about MQA. 8-]
 
Just like tossing a coin and constantly getting heads. A layman would assume that it was not random, but actually it is just unlikely
 
Thanks Jim for those links to your articles - I'll have a read soon

Somewhat related to this discussion about random & pseudo random noise, there is evidence that the auditory system summarizes the temporal details of sounds using time-averaged statistics.
There's a category of sounds termed "sound textures" such as the sounds of rain on a tin roof, applause, waterfalls, fire crackling, etc which we seem to recognise based on the statistical analysis of their sound spectrum. It may play a larger role than in just the perception of these categories of sound & be involved in our perception of other ubiquitous sound structures - such as our perception of room ambiance - in other words the background in which we perceive foreground sound?

Not only is noise a tricky area in audio electronics, it is also a tricky area in auditory perception.

"Summary statistics in auditory perception"
"Sound texture perception via statistics of the auditory periphery: evidence from sound synthesis"
 
You must log in or register to reply here.


advertisement


Back
Top