AD122-96 MX
Appendix I
13
dither have been explained by S. P. Lipshitz and J. Vanderkooy of the University of Waterloo.
Additional publications by M. Gerzon, P. G. Craven, J. R. Stuart and J. R. Wilson (all from the United
Kingdom) and J. Goodwin (from the U.S.A.) shed light on this complex subject.
Dithered noise shaping technology has been incorporated into a handful of hardware devices. While all
are based on the same concepts, some perform better than others. After simulating and listening to all
available public domain algorithms, Lavry Engineering came to some conclusions in forming a basis
for
α
coustic
β
it Correction™
. The principal conclusions are:
a. The practice of greatly amplifying low level signals to determine triangular flat PDF (probability
density function) dither reveals the effectiveness of distortion and noise modulation elimination. This
practice yields misleading results when testing unflattened dithers and/or noise shapers. It conflicts
directly with L. Fielder’s findings showing completely different threshold delectability curves for quiet
and loud levels. Noise shaping listening tests must be done at "reasonable" volume levels.
b. Given the above requirement, our listening tests concluded a strong preference for "triangle high
pass" dither (this dither is produced by simultaneously adding a new random number and subtracting
the previous value). Such dither is frequency-shaped to carry more high frequency energy (the energy
content at low frequencies is minimal).
c. Listening tests revealed a preference for smoothly varying noise-shaping curves. Peaks and notches
seem to irritate the listener (admittedly while turning the volume up). In addition, despite the
temptation to optimize the noise shaping curve to the average listener’s hearing threshold, given a
significant variation from listener to listener requires reasonable compromises in tailoring such a curve.
In other words, smooth the curve.
The improvements offered by dither and noise shaping vary with source material and final word length.
An A/B/X test at 16-bit level, requires a quiet environment and low level (loudness) audio. The listener
must resist the temptation to turn the volume up to unreasonable levels. The practice of truncating to
short word length (8-12 bits) should be avoided. The ideal noise-shaping curve may be irritating at
loud levels.
Lavry Engineering’s listening tests were based on test tones and repeating loops of quiet passages of
various material (mostly classical music) with flat amplifier response. Listening to test tones was
straightforward: we used the
Model AD122-96 MX
test tone generator mode switching the
α
coustic
β
it Correction™
on and off. The frequency and amplitude programmability was very useful.
Fig. 4 -
α
coustic
β
it Correction™
Fig. 5 - ABC HP-NS2
Fig. 6 - dither only
16 (top), 18, 20, and 22 bits
16 (top), 18, 20, and 22 bit
16 (top), 18, 20, and 22 bits
no input signal
-120 dB 1 kHz sine tone
-120 dB 1 kHz sine tone
α
coustic
β
it Correction™
may be used with words of lengths wider than 16 bits. Figure 4 shows the
noise floor of
α
coustic
β
it Correction™
(High Pass, NS2) without a signal at 16, 18, 20, and 22 bit
wordlengths. Note the curved noise-floor with lowest level in the ear’s most sensitive mid-range
region.
