dcs Purcell Upsampler
There are two sides to this coin. On the positive side, it definitely works, and provides a worthwhile improvement for your library of 16/44 CDs. On the negative side, it doesn't work in the way that most people say it does, so it does have limits, and it has a overall high expense once you count in the matching $12,000 Elgar (the only D/A convertor today that can use the Purcell to full advantage) plus a luxury CD transport worthy of this pair. If you don't want to spend the $30,000 this trio might cost, then you can get close to this sound with a well engineered integrated package for far less money, such as the Oracle CD player at $8950 total.
First, the good news. The Purcell, the consumer version of the dcs 972 professional upsampler, costs slightly less ($5000 vs. $7000), while offering the same processing power and sonic benefits. And the sonic benefits are considerable when applied to 16/44 CDs. We compared the sound of high quality (e.g. Chesky) 16/44 CDs, played straight through to the dcs Elgar D/A convertor, vs. the same CD signal being upsampled by the Purcell from 16/44 to its maximum capability, 24/192, and then being fed to the Elgar. The Purcell dramatically reduced the canned, closed in, glary, glazed, hard, slurred sonic qualities we heard from the 16/44 CD played straight. These negative sonic qualities are precisely the ones which most offend sensitive music lovers about CD sound, and which many have assumed were inextricably linked with the digital medium. The Purcell miraculously liberates 16/44 from the prison of these negative sonic qualities. The same CD heard through the Purcell sounded much more musically natural, more liquid, more relaxed. Subtle information contained in the recording became more audible, which also benefited stereo imaging and a sense of air around the instruments. Even the bass improved, sounding less boomy.
So now let's tackle the big question. How does the Purcell do this? The Purcell is a purely digital box; digital in and digital out. It takes in a digital signal at one sampling rate and one bit resolution, and outputs a digital signal at a sampling rate and bit resolution which you can select to be the same as the input or different. It can accommodate a wide range of input sampling rates (32 to 96 kHz). It can output a wide range of sampling rates (32 to 192 kHz) and bit resolutions (16 to 24 bits).
Presumably you will want to set up the Purcell's controls such that you use it to increase both the sampling rate and bit resolution of the input signal. For example, the most common consumer application at present would be to input a signal from a CD transport, which would have a 16 bit resolution and a 44 kHz sampling rate, and to increase the bit resolution and sampling rate to the maximum that your D/A convertor could profitably handle, say 24 bits and 96 kHz.
How does the Purcell produce a 24/96 output digital signal from a 16/44 input signal? Simple. Digital signals are simply numbers. So the Purcell uses a computer to calculate new output numbers (the output digital signal) from the input numbers (the input digital signal). Where does the higher sampling rate come from? Simple. The Purcell can be set up so its output clock runs faster than the input clock, thereby generating more output samples than input samples in any given time period. The digital values of the extra output samples could be mere repetitions of the digital value of the last input sample, or they could calculated by the computer to be interpolated values, between the last input sample and the next input sample. Where does the higher bit resolution come from? Simple. Any averaging, interpolating, or multiplying calculation can produce extra fractional detail (for example, the average of 5 and 6 is 5.5).
From the above, it might seem natural to assume that the Purcell improves the sound of your 16/44 CDs by increasing the bit resolution and sampling rate per se (after all, it is sold and promoted as an upsampler). And from this it might seem to naturally follow that the more one can increase bit resolution and sampling rate, the better. Indeed, that has been the tenor of some popular press about the Purcell's professional parent, the dcs 972 upsampler.
But these assumptions are wrong, or at least misleading. They suggest to you that we are once again in an audio horsepower numbers race (just as we once were with watts and then with THD). They tell you that any digital playback box which can give you higher numbers must sound better than one with lower numbers, because it is the higher numbers that are chiefly responsible for the better sound. They even suggest that a 24/192 upsampling of a 16/44 CD might sound better than a direct 24/96 DVD, simply because 192 is twice as high as 96. And you are already being told to junk your digital playback gear and buy expensive new gear, every time a new box comes along that can output a higher sampling rate and/or higher bit resolution. Tain't so.
Why not? Well, you see, virtually every humble integrated CD player already has a computer inside (as do separate D/A processors). And this computer already performs calculations on the input digital data coming from the 16/44 CD rotating inside. And this computer already increases the bit resolution to about 24 bits inside the CD player or separate D/A processor. And this computer already increases the sampling rate inside to 176.4 kHz (very close to 192 kHz) on many CD players with multibit DACs, and to an even higher 352.8 kHz on others (indeed all the way to 11,300 kHz on CD players with Bitstream or sigma delta DACs).
So, if virtually every humble integrated CD player already does all the increases, of bit resolution and sampling rate, that the Purcell does at its maximum setting, who needs to spend $5000 on the Purcell? And who needs to spend the $12,000 more, on the only D/A processor today (the dcs Elgar) that can accept the Purcell's maximum setting 24/192 output digital data stream?
To put the matter more precisely, if indeed the natural assumption were true, that the Purcell upsampler improves the sound of your 16/44 CDs by increasing the bit resolution and sampling rate per se, then the Purcell would have no advantage to offer over humble integrated CD players. That's why numerous articles in the high end audio press have wondered aloud in puzzlement: what's the true difference between common CD player oversampling and the Purcell's upsampling, and why does the latter provide sonic benefits that the former does not? To date, no one, including dcs' own redoubtable Mike Story, has been able to provide a satisfactory explanation.
Moreover, virtually all D/A convertors other than the $12,000 Elgar are restricted to processing no more than 20 bits of input data, and no more than a 96 kHz sampling rate. Thus, if you didn't want to spend $12,000 on an Elgar you would have to restrict the Purcell to less than its maximum output settings, say 20/96 instead of 24/192, to make the digital data stream acceptable to the input of other (less expensive) D/A convertors. But that would create a bottleneck which seems to make no sense. Why furnish these other D/A convertors with a digital data stream having poorer numbers (lower bit resolution and lower sampling rate) than they generate internally anyway, with their own internal computers? Some of these D/A convertors oversample to 352.8 kHz internally. Why feed them only 96 kHz? Who needs the Purcell computer to feed them poorer numbers than they already do internally on their own?
The Purcell seems caught in a curious self-contradiction. If indeed the natural assumption is correct, that the Purcell upsampler achieves its sonic benefits by upsampling per se, then it can't possibly achieve its sonic benefits! It can't achieve any benefits by upsampling because it is feeding poorer numbers to your D/A convertor than your D/A convertor already achieves internally by oversampling inside itself. But why then should one bother with the Purcell? And how then can one possibly explain the fact that the Purcell does indeed improve the sound from 16/44 CDs?
To find the answer, we must look to the wisdom of Sherlock Holmes. To paraphrase Sherlock, if an assumption fails to explains the facts, and worse yet runs counter to the observed facts, then that assumption, no matter how obvious or natural, must be wrong. In this case, the assumption that the Purcell works chiefly by increasing the bit resolution and sampling rate per se must be wrong. The Purcell might be an upsampler, but it does not magically create new musical information or achieve sonic improvement by upsampling per se.
What then is it that the Purcell does? When we evaluated the Purcell, we already knew that, even at its maximum settings, it was not delivering to a D/A convertor any significantly higher numbers (in terms of bit resolution and sampling rate) than the D/A convertor was already doing internally. So, with this foreknowledge in mind, here's how we evaluated the Purcell. First we naturally listened to the Purcell at its maximum settings of 24/192 output, and verified that it did indeed improve the sound of a 16/44 CD, compared to hearing the 16/44 CD direct (bypassing the Purcell's processing).
What did we do next? Did we try lower settings of sampling rate and /or bit resolution, as other evaluators have been tempted to do? No. We knew a priori that these would be inferior to what the D/A convertor was already doing internally, and there would be no point to feeding an inferior signal to the convertor. For example, we didn't bother trying to output merely 96 kHz from the Purcell, because this would be less than the 176.4 kHz or 192 kHz or 352.8 kHz that the D/A convertor would soon be upsampling the signal to anyway, and so would be a needless compromise bottleneck.
But we did try a different kind of comparison, to test the mettle of the Purcell and to verify our hunch as to how it really worked. The Purcell has user selectable noise shaping options. The maximum setting is 9th order noise shaping. After hearing that, we switched to the next lower setting, 3rd order noise shaping, while keeping the Purcell's output bit resolution and sampling rate at their maximum 24/192 settings. Bingo! The sound quality deteriorated, getting very close to that of the straight 16/44 CD sound. More importantly, the sound acquired precisely the negative qualities that we heard from the straight 16/44 CD, sounding glazed and canned.
This experiment confirmed what we had suspected: the Purcell achieves its sonic benefits primarily by aggressive, high order noise shaping, not by the mere act of increasing sampling rate, even with increased bit resolution. Even though the Purcell is called an upsampler, and even though the popular press is spawning a numbers race about upsampling, this experiment proved that there is scarcely any sonic benefit to the Purcell's upsampling per se, even at the maximum 192 kHz, and even when coupled with the maximum bit resolution increase to 24 bits. Indeed, even the Purcell's 3rd order noise shaping, which is already more aggressive than that used in the past (Philips used merely 2nd order noise shaping in its 14 bit CD player and later for even Bitstream with just 1 bit), scarcely provided any sonic improvement over direct 16/44, when coupled with upsampling to the maximum 192 kHz at 24 bits. It required the much more aggressive 9th order noise shaping to energize the Purcell's capabilities, and furnish a significant sonic improvement over 16/44 CD.
Hybrid Super PCM
What's this all about? What is noise shaping, and why should it help PCM digital? In an article written long ago, we had suggested that the ideal digital system would be a hybrid between PCM and sigma delta with noise shaping. Indeed, some of the best professional A/D recording convertors in recent years, based on Bob Adams' designs, have approached a hybrid concept (from a slightly different angle).
Here (in brief) is why a hybrid digital system could offer the best of both worlds. Multibit PCM digital captures the full resolution of music instant by instant, at each sampling point. So it can capture music's highest frequencies with the full bit resolution of the system (be it 16, 20, or 24 bits). But it does not afford any increased resolution beyond this to musical frequencies lower than the highest frequencies captured. For example, the crucial midrange frequencies, where so much music occurs and where human hearing is so sensitive, are not afforded any greater resolution than the trebles are.
On the other hand, sigma delta digital systems (chiefly 1 bit systems) chronically do not come even close to offering the true resolution needed to accurately capture music's highest frequencies (a 1 bit system would have to oversample at 32,000 times 44.1 kHz in order to offer the same true information content as a 16/44 PCM system, and no 1 bit system comes anywhere close to this). To counteract this shortcoming, sigma delta systems have relied heavily on noise shifting (noise shaping). Noise shifting or shaping is often spoken of as if it merely shifts noise out of the audio band. But it's actually more useful to speak of it in a different way. It's more useful to speak of this noise shaping process as an averaging process, which averages many 1 bit samples together to produce a single sample of music waveform. This averaging process reduces the very high noise of a 1 bit system, and it also reduces the very high distortion of a 1 bit system. But, most importantly, it also increases the effective bit resolution of the 1 bit system, thereby arriving at a pretty accurate representation of the musical waveform within the audible part of the spectrum.
However (and this is a big however), the efficacy of this bit resolution varies with frequency, and is (speaking oversimply) inversely proportional to the musical frequency being encoded and reproduced. The reason is simple. At lower musical frequencies, there are many, many digital 1 bit samples to average together, per cycle of music. The more digital data there is to average, the better the reduction can be of noise and distortion by the noise shaper, and the more accurately the system can approximate the true value of the musical waveform. Thus, the more digital data there is to average, the more the effective bit resolution can be increased. And, because there are more digital samples per musical cycle as the musical frequency gets lower, the bit resolution can be improved more for music's lower frequencies than for music's higher frequencies.
This accounts for the great weakness of most sigma delta (1 bit or low bit) systems. They don't have enough of a high oversampling margin above music's highest frequencies, so they don't have enough digital samples per musical cycle to average together at music's highest frequencies. Consequently their effective bit resolution is poor at music's highest frequencies, yielding trebles that are too soft, fuzzy, smeared, distorted, etc. (see our stories about the DSD digital system).
But this also reveals one of the great strengths of the sigma delta digital systems that employ noise shaping. At music's middle and lower frequencies they can provide dramatically improved bit resolution, thanks to their averaging of many, many digital samples to more accurately reconstruct the musical waveform. Indeed, with a well engineered combination of a high enough oversampling rate and an aggressive enough averaging algorithm, sigma delta systems can surpass most PCM systems in revealing subtle musical nuances at middle and low frequencies.
So PCM multibit excels at music's highest frequencies, while sigma delta with noise shaping averaging excels at music's middle and lower frequencies. Why not combine the strengths of the two digital systems into a hybrid system? Why not, indeed! In particular, why not start with a multibit PCM digital system and then bestow upon it the additional capability of noise shaping averaging. Learn a lesson from the noise shaping book that was required reading for sigma delta 1 bit systems, and apply it as an optional benefit to a multibit PCM system. The PCM system already has full required bit resolution (16, 20, or 24 bits) at music's highest frequencies, but it fails to improve upon this resolution for music's mid and lower frequencies. Why not improve upon PCM's resolution for music's mid and lower frequencies? Simply add noise shaping averaging, a technique borrowed from sigma delta, to PCM, and you can have super PCM, giving you full bit resolution at music's highest frequencies, and even higher (better, more accurate) bit resolution at lower frequencies. This would enable any given PCM multibit system to become even better at revealing the subtlest musical nuances.
That, ladies and gentlemen, is precisely what the Purcell does. And that's precisely how it sounds.
The Purcell applies noise shaping averaging to a multibit PCM input, and thereby arrives at an even more accurate approximation of music's middle and lower frequencies than are contained in the 16/44 digital PCM data coming from your transport playing a 16/44 CD. This might seem like voodoo magic, but it's not.
Sigma delta digital systems have only 1 bit of coarse resolution, which would sound awful (like hash that's either on or off). Yet, by adding noise shaping averaging, this 1 bit coarse hash can be formed into beautiful music. The noise shaping averaging essentially detects the statistical trend of the hashy on/off coarse 1 bit noise, and the computed trend of the coarse noise is the music signal, recovered with great accuracy (especially at middle and lower musical frequencies).
Likewise, instead of haughtily regarding the 16 bit resolution code coming off your CD as precise code accurately defining the music signal (perfect sound forever), we could humbly regard it as coarse noise that only crudely approximates the true nuances of the music signal (and this viewpoint has merit, since, as we showed by measurements in Hotlines 49 and following, human hearing can appreciate musical nuances represented by bit resolutions way down at the 20 bit level or lower). Once we have humbly understood that PCM's bit resolution can be improved, we can then use much the same techniques as sigma delta relies on, for noise shaping averaging, but applied instead to PCM. The noise shaping averaging applied to PCM would essentially detect the statistical trend of the coarse 16 bit code, and the computed trend of this coarse code is (just as for 1 bit sigma delta) the music signal, recovered with greater accuracy (especially at middle and lower musical frequencies).
If noise shaping averaging can work at all for a sigma delta digital system with just 1 coarse bit of resolution, improving the resolution so much beyond this 1 bit that this system can deliver real music having about 24 bits of effective resolution at middle and lower frequencies, then it's a cinch that noise shaping averaging can improve music when it has a much better input signal to work with, a signal that already has 16 bits of resolution instead of just 1 bit.
Why hasn't noise shaping averaging been applied previously to PCM, to improve it? Actually, it was already applied, in the earliest days of CD players. Philips (one of the pioneers of noise shaping averaging) found that their early DAC chips could not really live up to the 16 bits of resolution that
(Continued on page 22)