Intro
It is a well-known fact that if one looks at a speech waveform using
an oscilloscope it is quite commonly "lopsided." By this we mean that
one side of the waveform, say the top side, has a greater peak
amplitude than the other side. The degree of this asymmetry is highly
dependent on the voice of the individual involved. This paper will
explore this phenomena. We will outline possible problems which may
arise from it and methods for reducing or negating those problems.
Possible problems - entirely linear systems
In a high quality public address system where the speaker's
voice is simply being augmented by a power amplifier and
speaker there would be no problem if the volume levels are
such that the system is entirely linear. The lopsided
waveform would pass unscathed and indeed unnoticed.
If such a lopsided waveform were to be used to modulate an
AM transmitter in a noncompetitive situation, and if the
modulation is adjusted to be relatively low, then such
a waveform would offer no problem. It has been found
best in ordinary AM systems to set the polarity of
modulation such that the peaks with the greater amplitude
are "upward." This minimizes or even eliminates
distortion due to the typical envelope demodulator in the
typical radio receiver.
In a purely linear system asymmetric waveforms are not in
themselves a problem.
In systems using a fast-acting AGC loop
There was at one time a kind of contest between various
equipment manufacturers to see which one could develop
the fastest-acting AGC system to control the modulation
in a transmitter. The devices were commonly called
volume limiters or volume limiting amplifiers. They
generally reacted to an overload situation within a
millisecond or perhaps less. Following the overload
they would restore the gain to normal over a period of
time of perhaps a few seconds.
It was impressive to view the output of these devices on
an oscilloscope. There was no visible clipping or other
artifacts added to the waveform, just a (usually)
well-controlled modulation level. The winner in the war
of speed used a system which had a zero attack time. It
used a delay line to delay the audio signal while the
gain-controlling voltage was being generated.
But in every one of these units there was an annoying
tendency for the device to respond to signals which were
not the same as what the human ear responded to. Rephrased,
they were controlling modulation, not volume. Given a
transmitter which could handle the level-controlled
waveform properly, these units would control the modulation
levels nicely. But they did not maximize volume, which was
becoming an issue.
If a lopsided waveform were to be applied to one of these
units the peak with the greatest magnitude would cause
generation of AGC voltage. If the peaks (positive and
negative) could be made equal then less AGC voltage would
be generated and modulation would increase. This must be
done, however, in a manner which does not increase the
peak-to-peak value.
In systems involving clipping of transients
Research showed that if the AGC loop could have a reaction
time (attack time) of several milliseconds and a release
or recovery time of perhaps 100 to 200 milliseconds that
such an AGC loop would match the ear quite nicely. Such a
system would control modulation and it would match the
ear, allowing the volume to be maximized. There was a
drawback to such a scheme: the relatively
long attack time required that the basic AGC loop must be
followed by a clipper to catch the transients which escaped
it.
In a system with such a relatively long attack time an
asymmetrical applied waveform causes less AGC voltage
generation than in a short or fast attack time. However,
the signal from the AGC loop is then applied to a clipper.
If one peak is clipped more than the other we have a DC
(or at least subaudible/syllabic) component exiting from
the clipper. In an AM transmitter this appears as a form
of "carrier shift." In an FM transmitter this appears as a
center frequency shift. It ends up interfering with the
transmitter's AFC system. It is certainly disconcerting to
watch an analog frequency meter kick violently when such a
unit is used to control the modulation with an applied
asymmetric waveform.
To minimize this problem the waveform should be processed
in a manner such that prior to application to the clipper
the peaks are made symmetrical, but without increasing the
peak-to-peak value.
Let us look at ways to handle the problem.
A standardized waveform
First let us generate a standard waveform that can be
reproduced easily and will allow us to compare various
approaches to processing. The proposed signal here has
an approximate 10 dB (3:1 voltage-wise) positive to
negative ratio but it has no DC component. The
areas under the curve for the positive and for the
negative portions of the waveform are equal.
This signal has a fundamental frequency of 200 Hz
and is a believable replica of a steady speech
signal, having a similar spectrum. Such a modulating
waveform is illustrated here:
The waveform as shown would approximate the sound
"ohhh."
There is about 10 dB of asymmetry in this waveform
and there is no DC component; as can be seen, the
areas above and below the centerline are precisely
equal. At the end of this paper is a PSpice net
list to enable recreation of the waveform.
If used to modulate the transmitter directly
the upward peaks would require 10 dB more power
than the downward peaks. If the transmitter
could not handle this degree of asymmetry then
the modulation level would have to be reduced
until the positive peaks were in a linear region
and the negative peaks would be reduced in amplitude.
This is an inefficient use of transmitter capability.
Use a highpass filter
One way to accomplish the task is to apply the modulating
signal to a highpass filter. Indeed, this may be a part of
the speech-processing chain already. Such a highpass filter
would be doing double duty. It would be removing those
components that serve no purpose if transmitted, and in fact
may be causing mischief. The highpass filter can also make
the positive and negative peaks more nearly equal. If the
speech signal were applied to a highpass filter and then to
a clipper, there would be similar amounts of clipping applied
to the positive and negative peaks. There would be a
reduction in axis shift due to any subaudible components
generated by asymmetric clipping. And the ear normally
tolerates clipping of both modulating waveform peaks better
than clipping only one side of the waveform.
If that signal is applied to a 100 Hz highpass filter
we will obtain the signal shown in the following graphic:
This is the appearance of the waveform after it has
been passed through a 100 Hz highpass with a rolloff
rate of 18 dB per octave. It does make the positive
and the negative peaks have values closer to one
another, but at the expense of increasing
the value of the smaller peak. This is not our
object.
Further, the cutoff frequency of this highpass
filter would have to be readjusted for different
voices to maintain its effectiveness. This is a
conflict with its original intended purpose of
passing what we usually consider the "needed" speech
components.
Use an allpass network
Another way to process the speech signal is
to apply it to a network which rearranges the
relative phases of the signal to make it less
asymmetric (more symmetrical). Ideally this
network will have a flat amplitude response.
This can be done with a chain of simple circuits
using opamps, as shown in this schematic:
Note that the time-constants in the stages are
staggered. This allows the circuit to work regardless
of the frequency of the speech waveform fundamental.
Changing the frequency of the fundamental of our
computer-generated signal results in a fairly uniform
correction across the audio frequency band.
That is in fact one of the channels of a real-world
phasing network for SSB generation by the phasing
method.
The phase shift through that cascade of two
individual allpass networks is shown here:
The phase shift is from zero degrees at DC
through N times 180 degrees at infinity.
Here we have a total of 360 degrees of rotation
as we go through the audio spectrum (2 times 180).
The output of the allpass network is shown here:
The signal's peak-to-peak amplitude is not changed.
The magnitude of the higher-amplitude peak has been
reduced while at the same time the magnitude of the
lower-amplitude peak has been increased.
There is no axis shift and no subaudible components
have been added. The areas under the curve above
and below the zero axis are equal. Clipping of such
a waveform would cause a minimum of "mischief"
compared with clipping of the original asymmetric
waveform.
Failure
The use of an allpass network is not a cure-all.
If the applied waveform is symmetrical
top-to-bottom (i.e. has no even-order harmonic
content) then the allpass can actually
increase the signal's peak-to-peak
amplitude. Here we see an example of an
applied waveform with top-to-bottom symmetry:
This signal consists of a 200 Hz fundamental and
an equal amount of 600 Hz harmonic.
Now let us apply that signal to an allpass
network. The signal at the output of that
network can appear as shown here:
This signal still has top-to-bottom symmetry
but the peak amplitude has increased. This
can be seen by comparison of this waveform
with the previous. This is not necessarily
catastrophic but it illustrates a point:
If the waveform to be corrected has no
even-order components (unlikely in practice)
and so is symmetrical top-to-bottom in the first
place, then the use of an allpass network might
not be beneficial. This aspect of the allpass is
usually glossed over and is being pointed
out for completeness.
RMS-sensing AGC and clipping
If the modulation level in the transmitter is controlled by a
peak-sensing audio AGC unit ("limiter"), that AGC system will
respond to the peak with the highest instantaneous
magnitude. But if the modulation level is controlled by an RMS-sensing
AGC unit, asymmetry does not enter into the picture at all as regards
the AGC portion of the speech processing. But an RMS-sensing AGC
unit must be followed by a clipper to catch those waveform
excursions which escape the AGC unit. Be advised those excursions
will be of significance. But if the clipper operates on one side
of the waveform more than the other, a DC or subaudible component
will be developed by the clipper. This will normally cause trouble
in the modulator proper. We have a situation wherein the modulator
must be direct-coupled to properly handle the signal.
Here again the allpass can come to the rescue. Insert the
allpass between the RMS-sensing AGC block and the clipper.
By adding the allpass at this point the clipper will clip
symmetrically and no subaudible components will
be involved.
Suggestion
An allpass block should be placed in an appropriate
point in any speech processor used in a radio
transmitter:
It should be placed ahead of an AGC
circuit which is essentially instantaneous-acting
or
It should be placed ahead of a clipper if the
preceding AGC circuit
is low-acting or especially if it is RMS-sensing.
The allpass network is inexpensive and generally
insures lower distortion by virtue of less clipping
or at least symmetrical clipping (which sounds less
offensive to the human ear). Symmetrical clipping
always causes less "mischief" than does asymmetrical
clipping. Placement of the allpass block should
always be prior to the point where clipping might
occur or where it does occur.
History
The concept of using an allpass network to make an
audio signal more symmetrical is certainly not new.
The first commercial product was called the
SymmetraPeak and was placed on the market by
Kahn Communications. During the 70s an active
equivalent was installed in the circuitry inside
various audio "processors," including those produced
by Circuit Research Labs in Phoenix and by Orban
in San Francisco. The amateur radio fraternity
apparently has been relatively slow to pick up on
the idea. The only radio amateur writing on this
subject seen by ye scribe has been by Gary
Blau, W3AM.
W3AM Website
W3AM Allpass
The netlist
This netlist was written for the PSpice® Spice
simulator and is proposed as a generator of a
"standardized" speech waveform for allpass network
evaluation.
ALLPASS.cir
* 4 1 July 21 June 2001 J. Tonne
* Waveform generator; raw pulse is at 1;
* smoothed "voice" signal is at 100;
* output of highpass filter is at 200;
* output of allpass network is at 303 (for now).
* Raw pulse generator:
Vin 1 0 PULSE -.38 1 0 .0002 .001 .0008 .005
Rin 1 0 1G
* Response shaper to form voice signal:
E1 100 0 VALUE = { 1000 * ( V(3)-V(100) ) }
R1 1 2 10k
R2 2 3 10k
C1 2 100 .15u
C2 3 0 .003u
* Highpass driven by voice signal:
E2 200 0 VALUE = { 1000 * ( V(6)-V(200) ) }
R3 4 0 5.7158k ; These resistor values
R4 5 200 2.2469k ; make a Butterworth
R5 6 0 39.238k ; highpass filter
C3 100 4 .2u
C4 4 5 .2u ; .2uF yields -3 @ 100 Hz
C5 5 6 .2u ; with resistors shown
* Allpass driven by voice signal:
R6 100 301 10k
R7 301 303 10k
R8 302 0 10k
C6 100 302 .1u
E3 303 0 VALUE = { 1000 * ( V(302)-V(301) ) }
.TRAN 10u .06 0 10u
.PROBE
.END
Questions? Comments?
the author for further information, suggestions,
or to comment on this material.
This is a first pass at a downloadable/browsable
tutorial. Your feedback and comments are vital to
the effort, and are welcome.
|