Last issue we saw that nostalgia for analog may be misplaced, what with its frequency response and dynamic range problems and constant vigilance needed to get good performance. However, some whove heard me talk on this topic, and those of you who read the last issue, now know that the other shoe is about to drop, too digitals got problems as well. Bits Are Bits, No?
When I set out to make a set of perfect test signals available on CD,1 I had heard, as you doubtlessly have, that CDs dont sound like the masters that made them. What was going on? Some subtle jitter effects? Maybe the bits were not being copied directly, and the disc manufacturing was subjecting them to some changes? So what I did was to accompany the test disc tracks throughout production from the earliest stage to the replicated discs with a bit transparency test signal Prism Sound DAS90s confidence check. This check exercises all of the bits in a repetitive pattern that can be played back into a DAS90 receiver, and, with the correct matching pattern stored, compares the recorded with the exemplary pattern, and reports any differences. I fully expected to find mismatched bits between source master and reproduced CD the source of the sonic problem reported.
Im here to tell you that disc replication was just about the only step in the chain that was in fact bit transparent. Hidden features and even apparent mistakes in workstations caused trouble, as did DATs and CD-Rs. One workstation problem only went away when the host motherboard was changed, not the DAW hardware or software! It seems anything that touches the signal can get into the act, and does. Of course, you may wish the workstation to change the sound you may be equalizing it, for instance. Of course, there are going to be differences then, and probably differences among various workstations set to the same settings, as their design engineers define shapes of equalization curves differently, for instance.2
Thats always going to happen, but what I am speaking about here is more fundamental: a DAT machine that covers up dropouts sufficiently well for transparency on music, but for which the confidence test shows the bits are not the same; a CD-R recorder that apparently contains an input sample rate converter, working even when no conversion is being made; a workstation that contains a dc filter (very low-frequency high pass) and applies it to everything in order to prevent A/D offset voltages from causing a click at edits. The strangest of all was the host motherboard problem, probably the result of some memory mismanagement, I presume. These are all situations, and there are many of them, where we as audio engineers expect bit transparency, but apparently dont get it all that often.
The answer to this is to test the path you are using with a confidence check test signal or to compare, bit for bit, the output file from a process to its input file, a tedious process. This is what we are doing when we listen throughout a DVD-A production test to a few of the channels at a time. This may be the digital era, but were still doing a lot of hand work. Its crazy because there oughta be a machine
Hats off to MLP, which, once you get content into it, contains a running self check of bit transparency so that decoders will simply stop playing if the stream is not bit transparent. We can only wish that professional digital audio equipment will do the same thing some day. It is part of the squandered potential for digital audio that this isnt so yet. Remember when digital equaled good? When the IEC Audio, Video, and Audio/Visual Systems committee met in Athens some years ago and the question was asked whether digital in and of itself meant good, there was no one who would agree with this sentiment! Fundamental Conversion
The type of conversion to use is a hot debate. You couldnt have any two statements in audio further apart than the titles of AES preprints 5395 and 5396, covered in this months Relevant Research column: Why 1-Bit Sigma-Delta Conversion is Unsuitable for High-Quality Applications, and Why Direct Stream Digital (DSD) is the Best Choice as a Digital Audio Format. These emphasize why there is a process to preprint AES papers given at conventions, then submit them to peer review before publication in the Journal. While the debate rages, were caught in a potential dual-standard world. For those of us in audio for picture, this debate is moot, because it doesnt affect us: PCM is the only game in town as an originating mechanism, and low-bit-rate coded audio is virtually the only release medium. The debate there is between AC-3 and DTS, but the terms of that debate are pretty well known. Well keep up with the ongoing debate on DSD vs. PCM in future articles. DSD advertises itself as a 1-bit system, sampled at 64 x 44.1 kHz, that trades off the required amplitude accuracy of the multi-bit converter for supreme accuracy in the time domain, requiring extremely pure clocks. A problem for 1-bit conversion is the presence of idle tones in the noise floor and whether these are solvable or not is one of the main debates between the two preprints cited above. These can be audible and ruin things, and it has now been announced by one proponent that DSD uses multi-bit for a part of the processing to avoid these idle tones, a known fault of 1-bit conversion. So much for the marketing story, I guess. But the jury is out on the overall approach, even if it is more complicated than the simplified marketing story.
Multi-bit Linear PCM is well established and understood. With increases to sample rate and word length of DVD-A to beyond any practical hearing limits, by a great deal, and with the application of the required correct dither, it can be shown to be theoretically transparent when the steps of the converter are all equal. Yet there are deviations from theory in any practical product, ranging from the benign to the fatal. Some of the worst things weve seen are stuck bits. A test for this is on Disc 2 of the Hollywood Edge Test Disc series. It is a low-level sine wave tone accompanied by 2-minute infrasonic ramp. The tone should simply sound constant throughout the two minutes, and on many converters it does. On some converters, however, this signal goes so far as to shut off the output for a part of the time more than a few seconds. So the presence of a strong infrasonic signal might make this converter cut in and out. Others show grittiness during part of the infrasonic ramp, demonstrating that the converter exhibits differential nonlinearity perhaps up to and including nonmonotonic performance (where the converters steps are not all ascending, but some even reverse the trend and decrease in value). Another fault is shown by simply running a computer-generated, continuously swept sine wave with proper dither at 20 dBFS through a DAC. Weve heard birdies, added tones descending, then ascending again, as the desired tone ascends. This was on a reference player at one of the Hollywood production studios, one that is used for playback quality control.
With the swept tone at 0 dBFS, another set of factors comes into play ripple. Ripple in the output DAC filter, even of ±0.0005 dB, makes the perfectly coded swept sine wave alternately clip and not clip. This would not be so bad (because who works to these tolerances with program material) if it were not for the fact that some DACs wrap their output level when they clip. This means that they jump from the positive output extreme to the negative, and back again, when clipped positively, an apparent result of mathematical errors in their anti-imaging filters.
You can test for this and several other things by looking at a fully 0 dBFS peak coded square wave. While it should be perfectly square, inevitably there has to be an overshoot and usually pre-ring, plus post-ringing after the transient. The overshoot pushes the output level requirement beyond 0 dBFS into never-never land. By looking at the square wave with level recorded clear up to 0 dBFS, then descending in small steps to a few dB below overload, one can examine the nature of the filtering used, and any mathematical problems causing wrapping or disturbances even below the maximum output level. Pick Your Poison
Ive tried to be fair and evenhanded to both analog and digital camps in these two articles, and maybe made myself into an equal opportunity offender to some. There just isnt one approach that simplifies things down to this is good, that bad. Both analog and digital have their Achilles heels and bright points, but it has to be said that the modern world is increasingly digital. Thats an answer to Shakespeares question in Macbeth: If you can look into the seeds of time, and say which grain will grow and which will not, speak then to me.