The Wayback Machine - https://web.archive.org/web/20030208112337/http://www.

taddeo-l…

Taddeo Loudspeaker Company

2604 Elmwood Ave. Suite 105 Rochester, NY 14618
877. 625. 4894

Important News!

Technical Scientific The Future of

Reviews Links
Explanation Evidence Digital

Scientific Evidence, the Antidote Two is not snake oil!

This excerpt is from a paper on "fiber Optic Transmission and Professional Audio" by Dan
Talbot, published in the May '94 Vol. #42 edition of the Journal of the Audio Engineering
Society. The excerpt points out the flaw in digital audio that the Digital Antidote
addresses. This is the most significant flaw in digital audio. The author also points out
"that this fact is missed by some marketeers." More properly, one could say that this flaw
had gone almost totally unrecognized by the audio industry.

"In CD and DAT audio a sampling rate of near 44 kHz is used. This is about the minimum
allowable to process dc to 20 kHz audio without aliasing. To prevent untrasonic whistles
from being converted to audible interference, an antilias filter must be used to limit the
frequency range to dc to 20 kHz. Since 20 kHz audio is close to the 22 kHz band edge, a
very steep filter is required. The steepness of thes filter creates ringing on transients
whether or not this filter is linear phase (a fact missed by some marketeers.) The ringing
predominates at a frequency of 20 -- 22 kHz. This imparts "tonality" or "Johnny one-note"
quality to cymbal crashes and voice sibilance. If CD audio bandwidths were increased, less
playing time would be available on the disk (assuming all other factors remain constant.)
For this reason, the 44 kHz sampling rate represents a compromise. Many engineers have
worked to ease the realization of the 20 kHz antialias filter by using oversampling.
Oversampling also eases the problem of realizing linear phase. However, the basic problem
remains: a 22 kHz brick wall pass band, and this brick wall is the root of the ringing
problem."

The following column by Peter Van Willensward appeared in the Dec. 1999 issue of
Stereophile. It validates what we have been saying since 1993, when we first put the
passive version on the market.

The Netherlands
Peter van Willenswaard

Do we really need 96kHz (for higher) sampling rates? Those who have heard them,
including me, say they sound better—expecially the 192kHz and DSD varieties. But
scientific proof that humans can actually hear above the 20kHz barrier broken by these
high-sampling schemes is scarce and debatable.
Last year, recording and mastering engineer Bob Katz, best known for his purist
recordings on the Chesky label, set up an interesting test to find out if the audio
frequencies above 20kHz were responsible for the improvements heard with the higher
sampling rates. He recorded a 96kHz-sampled piece of music on his Sonic Solutions
digital audio workstation, then ran that through a custom-designed digital 20kHz
brickwall low-pass filter to mimic the standard CD audio bandwidth. However, the filter
did not downsample the music to 44.1kHz; its output remained at the original 96kHz
sample frequency. The advantage of this was that the D/A converter used in the
comparison would always operate under the same conditions; the only variable was
whether or not the audio cut off at 20kHz.

To Katz’s surprise, he could hear no difference between the original and filtered versions.
He reported his findings to a professional audio discussion group on the Web and
created a long-running thread.

Intrigued, my Dutch colleague Eelco Grimm and I decided to try to set up such a test
ourselves. We had long been of opposite positions on the subject of "audible above
20kHz," his stance being that, if properly implemented, 20kHz of bandwidth for digital
audio should be enough. Still, we agreed that the high-sampling demos we’d heard
indeed has sounded better. But it took a while to get all the necessary equipment in one
place at the right time. Meanwhile, by the end of last year, at the Tonmeister Tagung in
Frankfurt, a group of German recording engineers reported a similar experiment, this one
carried out with a Spectral Design AudioCube digital recording system. Result: They, too,
heard no difference between filtered and unfiltered versions.

Finally, April 1999, we had been able to bring together a range of equipment that would
allow us to make our own comparisons with three different systems: 1) Sonic Solutions
digital audio workstation with Katz filter, the "Katz Test"; 2) AudioCube, using their
DoubleRate SRC (sample-rate converter), German variant; 3) dCS 972 Sample Rate
Converter, with Augan OMX-24 hard disk as intermediate storage device (fig.1).

It is illustrative to note that both

the Katz and the AudioCube
tests use brickwall filters,
whereas the dCS features a
slightly "softer" rolloff (dCS
filter #2). The audio data was
processed twice by the filters in
the dCS and AudioCube: once
down to 48kHz, and again up to
96kHz. The data was
processed by the specially
designed Katz filter only once
(it couldn’t be used in real time,
by the way, as it ran as software
on a computer), but had the
disadvantage of being closer to
the audioband than the 24kHz brickwall filters of the other two.

As a precaution, to exclude differences in jitter between the various filter paths that
might unwontedly influence sound quality, we copied all originals and all filtered results
to the same Nagra-D, a 24-bit/96kHzdigital tape recorder of known low-jitter
performance. "Originals" is plural, as we listened to four different recordings, all 24/96.
All listening was done via the same dCS 954 DAC (in 24/96 mode) into high-quality
active two-way studio loudspeakers made by Heynis, a Dutch company. Location was
the mastering room of Bert van der Wolf (Kompas CD Multimedia), who is also the Dutch
dCS distributor and one of the pioneers of high sampling rates in Holland. A fourth
person taking part in the listening test was Onno Scholtze, a freelance recording
engineer formerly with Philips Classics. Onno, Bert, Eelco, and I are all audiophiles, and
none of us shies away from controversial subjects (which takes courage if you operate in
the professional audio sector, I assure you). The listening took an entire day. I don’t have
room here to relate all the details, so I’ll present just our conclusions.

First, all four of us heard differences between the original and the various low-pass
filtered versions. The original sounded best (no surprise, maybe) to all of us, with the
dCS route coming in a unanimous second. When it came to the Katz filter and the
AudioCube SRC, performances depended on the listener’s taste.

The dCS proved to have the best sounding filter, but was the difference from the original
due to the filter, or to the use of the dCS 972 itself? Bert had looked into that earlier:
When using the 972 to go from 96kHz to 44.1kHz and back up to 96kHz, he had heard no
difference.

Bob Katz hadn’t listened from a Nagra -D as a source, but directly from the Sonic. As the
German Tonmeisters had used their AudioCube as their source, we decided to compare
original to filtered versions in those ways too. The differences now almost vanished—
apparently, the higher jitter of these latter recording systems effectively masked the
influence of the brickwall filtering. Bob Katz had been right when he reported not to have
heard a difference—when using the Sonic, neither did we. We had just been lucky to
choose the low-jitter Nagra-D as a common source in our three-system comparison.

When Eelco reported our findings to the Internet discussion group, all hell broke loose.
They said we should have been A/B switching instead of listening to a few minutes of an
original, then a few minutes of a filtered version. Some would take our findings seriously
only if we would repeat it under (double-) blind conditions.

So, after weeks of e-battering, Eelco and Bert reluctantly planned a new session to see
what they could do. They first verified that the previously heard differences were still
audible via our standard nonblind method. They were. Then they started A/B switching
between synchronized files—using the Augan hard-disk recorder, also a very low-jitter
device—and confirmed what they already knew: A/B allows an evaluation of just one
difference aspect; if the difference is subtle, it will be heard only during the first few
switches, after which A/B switching becomes very frustrating. Nevertheless, they also
set up blind tests for each other ("difference or not") in which each A/B pair could be
listened to as often as requested. After each listener had worked his way through four
different music fragments, they gave up, exhausted, and counted their scores: 50% pure
chance.

Then Bert came up with an idea: What would happen if the four variations (original, Katz,
dCS, AudioCube) were recored in random order and numbered 1, 2, 3, 4? That, too, would
be a blind test. With Bert absent from the room, Eelco prepared a set of files in the
Augan, numbered them, noted for himself which was which, and went home. Bert
couldn’t wait, and took the test the same evening. Before he started listening—this is
crucial—he deleted the reference files in order to come as close to the sighted procedure
we followed earlier. Then he compared files 1, 2, 3, and 4, noting his preferences: best,
less good, poorer, poorest. When he phoned Eelco with his results, he had given the
various tracks exactly the same ranking as he had during the sighted test—meaning that
he had a score of 100%. This could have been chance, of course, but it wasn’t the first
time he’d performed a test this way. He had used the same method several times before
when evaluating the new sets of digital filters dCS sent him now and then, and had
ranked those in the correct order as well—every time.
All of this, too, was reported back to the Internet discussion group. The evidence was
almost unanimously accepted.

So, it seems, we discovered a blind test that works: a ranking or preference test. The
original may be included as one of the items, but its identity must not be disclosed to the
listener, nor must it be seperately available as a reference, lest one fall into the pits of
A/B testing and confuse oneself with mind-cracking endeavors to try to hear differences
where none exist—which is what happens when you compare original with original. The
interesting thing is that this blind preference test has a great similarity to what to what
we audiophiles do in our sighted experiments: We will change something, listen, say
"This is better." then change something else, deliver the next verdict, etc. If we repeat the
evaluation a week later, we usually end up with the same conclusions and rankings.

I’ve concentrated here on test methods. But what about the opening question: Do we
need higher sampling rates? Well, we still don’t know for sure. It’s evident now that
brickwall filters, and even "softer" digital cutoff filters near the audioband, can be heard.
But are we hearing the lack of ultrasonic frequencies, or are we hearing the filtering
action?

Maybe we’re stuck in a paradox: We can’t hear frequencies above 20kHz, so we can
throw them out—but if we do that, we’ll hear the filter, so we can’t.
All material herein is copyright © 1999-2002 Taddeo Loudspeaker Co. and may not be duplicated in whole
or in part without express written permission. All rights reserved. This technology is protected by U.S.
patent #5436882

Return to the Taddeo Home Page