Upsampling or Oversampling?
I first heard this at HI-FI '98 in Los Angeles, where Steven Lee of Canorus, the then distributor of Nagra and dCS, was using a professional dCS 972 sample-rate converter to upsample 44.1kHz audio data, first to 96kHz, then to 192kHz. With each change, there was an unambiguous improvement in sound quality. When first Jonathan Scull, and then I, tried a dCS 972 at home in our own systems, we were again impressed by the improvement—so much so that I made the dCS 972 my "Editor's Choice" for 1998.
However, as my measurements accompanying Jonathan's February 1999 review of the 972 revealed, when the unit upsampled CD audio data, it didn't add high-frequency information above the CD's original Nyquist limit of 22.05kHz. And while the dCS unit can be set to add eight least-significant bits of random dither noise below the 16th bit, it didn't add any information that was below the CD's capability to preserve.
While I was content to believe the evidence of my own ears—I even bought a 972—my experience with the upsampler left a major question unresolved in my mind, one that reader Don Hanlon repeats in this issue's "Letters" (p.9): How did what the upsampling dCS do differ from what the oversampling digital low-pass filters universally used in CD players and D/A processors since the mid-1980s have been doing all along?
Some background: The raw spectrum of a sampled digital audio datastream contains what are called "images" of the baseband data above and below each multiple of the sample rate (footnote 1). Thus, in a 44.1kHz system, the baseband spectrum extends from 0Hz to 22.05kHz, the first image is inverted and extends from 44.1kHz down to 22.05kHz, the second image extends from 44.1kHz to 66.15kHz, and so on, ad infinitum. Drastic low-pass filtering is required to eliminate these images on D/A conversion, leaving just the audioband content in the re-created analog signal.
For its first CD players, Philips realized that by adding three zero-valued samples in between each valid audio sample, the effective sample rate would be increased to 176.4kHz, and the unwanted images, instead of lying 22.05kHz on either side of 44.1kHz and its multiples, would now lie at either side of 176.4kHz and its multiples. A more gentle, better-behaved, and more consistent analog filter could therefore be used to eliminate the images. The output level would be now one quarter of what it had been, but that could easily compensated for in the analog output amplifier.
In the years since then, the audio industry has settled on an 8x-oversampling ratio, the 44.1kHz CD data being converted to a 352.8kHz datastream before D/A conversion by such popular chips as the Pacific Microsonics PMD-100. Companies like Theta, Wadia, and Krell have used proprietary filter algorithms, but these, too, oversampled the data.
So while I strongly suspected that the improvement I heard with the dCS 972 was simply due to its using a different oversampling filter, along with the benefit of better downstream DAC behavior when fed a 24-bit rather than a 16-bit signal—as I described in my January 1996 review of the Meridian 518—I wasn't sufficiently sure to spill ink on the subject.
But, perhaps partly as a result of this magazine's positive coverage of the dCS upsampler, a veritable slew of products has appeared offering that "96kHz" magic bullet. Like the Bel Canto DAC 1 reviewed by Robert Deutsch in this issue (p.143), or the MSB LinkDAC III chosen by Stereophile's scribes as our "Budget Component of 2000" (p.69), many of these products use Crystal's new CS8420 sample-rate converter chip to produce a high-sample-rate datastream from CD data. Others, such as the dCS 972, use a digital filter with several choices of topology and noise-shaping behavior.
Now I am sure. It is important to remember three things about all of these products: 1) other than making active the lowest 8 bits of a 24-bit word, no new audio information is created by any of these products; 2) as susceptibility to word-clock jitter increases with sampling frequency, it is always possible that upsampling audio data can make things worse, not better; and 3) no matter how good these upsampling products can sound—and the dCS, Bel Canto, and MSB products indeed sound excellent—there is no conceptual difference between them and traditional CD playback systems. I am now convinced that the sonic differences we have heard and reported on are due to the different choices in digital filters made by the designers of these products with respect to the number of taps, passband ripple, and stopband rejection (footnote 2), and to changes in the jitter performance.
We will have more on this knotty subject in our January 2001 issue, when Jonathan Scull and Bob Deutsch, respectively, review the dCS Purcell and the Perpetual Technologies PA 1 digital/digital processors. In the meantime, don't buy a digital product because it has "24/96" emblazoned on its front panel. Buy it because it makes your CDs sound great. And if it can accept real hi-rez data from an SACD or DVD-A transport—or can be upgraded to do so, given the probability that these datastreams will be encrypted—that's a bonus.
Footnote 1: An excellent essay by Jon Herron on the subject of oversampling digital filters can be found on the Madrigal website.
Footnote 2: A good examination of digital low-pass filter behavior, "Effects.pdf," can be downloaded from the dCS website.