[net.audio] Phase Distortion in Anti-Alias Filters- Synopsis

shauns@vice.UUCP (Shaun Simpkins) (01/22/85)

Hi, Gang... I'm back with the goods on reconstruction filter phase distortion
like I said I'd be. This is a synopsis of a report given at the 1983 AES
convention on the audibility of phase distortion of different reconstruction
filter characteristics.  Authors: D. Preis of Tufts University and P.J. Bloom
of Polytechnic of Central London.  Source: JAES, Vol 32, No. 11, Nov. 1984.

		     ABSTRACT

Experimental results on the audibility of phase distortion produced by minimum
phase 4-KHz and 15-KHz anti-alias filters are reported.  Numerous impulse
response pairs with identical spectral magnitudes but linear-phase and
minimum-phase responses were compared by five test subjects who listened
diotically (same signal at both ears) on earphones.  Group-delay distortion
was doubled, progressively, until at least 67% mean correct discrimination was
attained.  At 4KHz the phase distortion introduced by a cascade of two eighth-
order Butterworth filter pairs was audible, as was that from only a single pair
of seventh-order elliptic filters.  At 15KHz the cascade of up to 4 pairs of
seventh-order elliptic filters introducted no perceptible effects.  Experimental
details and several representative filter frequency and transient responses are
included.

The test setup was interesting; an analog low-pass filter of the desired type
was impulse-tested, spectrum analyzed, and the resultant characteristic
simulated digitally.  This was then convolved with itself to produce a group-
delayed response of two filters in cascade, and autocorrelated with itself to
produce a linear phase two-filter cascade response with identical amplitude
characteristics to the group-delayed pair.  Cascading was simulated by repeated
autocorrelation or convolution.  The test signal used was a 10us wide pulse
(no rep rate given - I assume it was 1KHz), digitally generated.  The output
signal was converted to analog by a 16-bit D/A and a Barr+Stroud EF3-04 20KHz
reconstruction filter.  Each pair was presented in random sequence (AA AB BA BB)
to the user who was then asked to identify the phase-distorted response from
the linear phase response.

Obviously, we can nitpick about the seemingly circular method of test - a test
setup with a reconstruction filter in it, with a cutoff frequency quite close
to those filters tested, is being used to test the audibilty of reconstruction
filter characteristics.  However, the LPF is a constant in all tests and should
therefore have an equal effect on all decisions.  
I still would like to know what the response characteristic of the LPF used was
event though the results of the test say that this would be just an idle
curiosity.

From this, the authors conclude that, under the conditions of the test, the ear
is far more sensitive to phase distortion in the midrange than the treble,
where most of the phase errors in CD reconstruction filters occur.  The use of
filter pairs simulates the band-limiting filter at the digital recorder inputs
in cascade with the CD player's reconstruction filter.  It isn't likely that
more than 4 of these pairs would be passed in the creation of a CD (including
repeated conversion for mixdown if a digital mixer is not available).  The
authors state their own reservations about the ultimate validity of their test,
which I let stand on its own:

	No attempt has been made to determine the detailed dependence of
	the perceptual threshold on frequency, or the peak and width of
	the group delay characteristic, or signal intensity, or signal
	polarity.  Other test signals, such as speech or music, were not
	used.  No other methods of irradiation, such as loudspeakers in
	nonreverberant or reverberant environments, were tried.  Most likely
	these are all important factors, which would produce measurable
	shifts in perceptual thresholds, and they are, therefore, certainly
	worth investigating.

Comments, netters, please - but before you comment blindly, read the full
report.

The wandering squash,
-- 
				Shaun Simpkins

uucp:	{ucbvax,decvax,chico,pur-ee,cbosg,ihnss}!teklabs!tekcad!vice!shauns
CSnet:	shauns@tek
ARPAnet:shauns.tek@rand-relay