jas@proteon.com (John A. Shriver) (04/22/91)
The following letter was printed in the April 1991 issue of High-Fi News & Record Review (an English publication). It is a response by a digital audio designer to a rather dreary article of the "here it is, it's digital, by definition it's more perfect than before" that had run a few months before in the same magazine. I find it a good debunking of some of these myths that the naive cling to, and find it frustrating only in that it does not name names. (I presume the author doesn't want to give away all his trade tricks and secrets, nor annoy his vendors.) Actually, those programmers among us who have programmed some of the wonderful interface chips like the Zilog Z8530, the Intel 82586, the National Seminconductor DP8390 (with their multi-page errata sheets, if you can get copies of them), know just how perfect this world's chip designers are. Now the same folks are designing noise shaping chips with roundoff errors, and hoping we won't check their math... Herewith, with apologies for my typing, proofreading, and any accidental undoings of the Queen's English (checked with spell -b): Dear Sir, I hesitated before writing this letter because I don't really want to criticise one of your contributors. However John Watkinson's `Simple Sampling' in the January and February editions deserve some comment. A certain amount of digital audio theory has been described and used to justify the claim that `well engineered' noise-shaping CD players reach the quality limits determined by the format and are essentially transparent. I'm touched by the simple trust that John has in digital technology; by contrast I'm becoming a hardened cynic. Just where are all these `well engineered' products? In the real world CD players don't seem to meet the Watkinson definition of `well engineered' and in my own occasional lectures I demonstrate quite audible (and measurable) problems which can arise in the performance of players from some of the most reputable manufacturers. I've designed many, many players -- always to the satisfactions of my clients yet there hasn't been a one that I couldn't go back to and improve. I don't think that I'm particularly incompetent it's just some of the failure mechanisms in digital audio are particularly insidious. To explain further would take up as much space as John's original article but readers should be aware that one of the problems of digital processing (such as is found in digital filters) is that is just one sample has an error, then the next, say, 512 samples will also be in error. Perhaps the biggest fallacy is that in digital logic a `1' is a `1' and is `0' is a `0' and, to quote: `...and therefore the digital circuitry of a well engineered CD player has no quality. Since this is a fact, we must look elsewhere to find the factors which determine the quality in a digital audio device...' Well, I'm afraid that this `fact' belongs in a different category that if you punch a brick wall you will invariably hurt your hand. For a start, one the data have been recompiled they are outside the of the error correction system and totally unprotected from degradation and corruption. And corruption can and does occur, as any designer of high-speed logic will confirm. Space again prevents a detailed discussion but as an interesting aside readers might like to turn to page 106 in the January edition of Stereophile magazine where they will find part of an interview with Richard Cabot, one of the designers of the Audio Precision System One; the world's finest piece of audio measuring equipment used by designers such as myself and HFN/RR's technical reviewer Martin Colloms. Cabot is a man who is really interested in measured performance and by his own admission is not a `Golden Ear', yet he was amazed to find that if he `improved' the quality of the square waves between the digital outputs of a digital filter and the digital inputs of a converter then the measured linearity improved by 6dB at low levels -- a dramatic improvement. Yet these are the same `1s' and `0s' that allegedly have no influence. Unfortunately the author's comparison to computers running word processing software isn't a very suitable analogy. There seems little doubt that noise-shaping `bitstream' technology is where the future lies. It was first conceived in the early 1970s but only the availability of low cost 20-50MHz chip technology has allowed it to become a commercial reality. Even so John Watkinson is deluding himself if he believes that the current incarnations of bitstream have got it all right. What about the audible `idle pattern' signals one manufacturer introduced; or the Japanese chip that has possible instability conditions; or the only party-explained distortion characteristics from another manufacturer that can only be minimised by `frigging' the analogue circuit design; or the audible modulation noise effects exhibited by one of the most popular noise-shaping chips; or the abysmal noise performance of some of the early chips that could only be hidden by quoting `A weighted' noise figures whilst omitting to mention that same fact. I could go on and on but the truth is that noise shaping today still has its own bugs which have to be worked through. I know, I'm one of the poor souls who has to get the circuits to work even though the maker's specification sheets turn out to be wildly optimistic. Yet for the armchair theoreticians even a study of the mathematics will show you that you can't get everything for nothing; there are still some inherent problems associated with noise shaping technology that have to be solved. Experience has forced me to have to assume that some of the chip designers don't know the problems are there, or perhaps it is simpler to ingnore them and hope that no-one notices. `Bitstream' may be the fashionable technology in the marketplace but it is worth remembering that the audio D/A converter that offers the best standards of performance (on any measured parameter) is a conventional multibit device. Finally we come to a test which will show that any two well-engineered noise-shaping CD players will sound the same in that their synchronised outputs will be identical to -90dB. I'm sorry, but what a load of twaddle! Even if we assume (and boy are we talking of a seriously unjustifiable Mother and Father of assumptions) that the two analogue stages have identical performance (other than of gain), then there are still many reasons in the digital domain that the test will not work. I'll leave you with just a couple to ponder. There is no reason to assume that the two noise-shaping systems are identical; they may be of different orders or have different feedback coefficients in their `feedback' loops. The end result will be a different distribution of noise in the upper-mid to high frequency regions. A few back-of- envelope calculations will show there to be differences well within the 90dB window. But even with identical noise-shaping DACs it must be remembered that their noise energy and spectra are only identical on an averaged basis. The same noise should be random in nature which by definition means that it will not be the same for both players at any given instant. Is there any reason to assume that the oversampling FIR filters will be identical? This hasn't happened yet, so I see no reason to believe it ever will. Firstly the frequency responses in the region of roll-off and above could well be different in the 15 to 20KHz region, perhaps by only 0.2 or so dB but that's enough. What about the in-band response? Many digital filters have virtually zero in-band ripple, but not all because it just one of the factors that a designer can trade off in order to optimise the overall filter response. The ripple is typically of the order of plus-minus 0.05dB. Indeed almost nothing, but when a 0.1dB difference in response between the two players will result in a spectrum of signals only 36 or so dB down and at the frequencies corresponding to the ripples. Obviously this would be enough for the test to fail, by definition, the player would cease to be `well engineered". If that proved to be the sole basis for rejecting a CD player than I suggest that the air would contain a few choice phrases borrowed from by fellow Yorkshireman Geoff Boycott. Stan Curtis, Consultant Designer, St Ives, Cambs (Copyright 1991 Link House Magazines Ltd) [Is this fair use?]
igb@fulcrum.bt.co.uk (Ian G Batten) (04/23/91)
In article <11263@uwm.edu> jas@proteon.com (John A. Shriver) writes: > > Stan Curtis, Consultant Designer, St Ives, Cambs I saw the letter that has been quoted, and broadly agreed with it. Curtis designed the Cambridge CD1, CD2 and CD3 with their 16x16 layout as well as doing the British modifications to Rotel's 16x4 CD players. He also did most of the Cambridge Audio amplifiers of the late seventies and eighties. I hope that clues you in on which machines he's talking about. ian