Report No. 3947
Analysis of Recorded Sounds Relating to the Assassination of President John F. Kennedy
James E. Barger,
Scott P. Robinson, Edward C. Schmidt, and Jared J. Wolf
January 1979
Prepared for:
Select Committee on Assassinations
Bolt Beranek and Newman Inc.
50 Moulton Street
Cambridge, MA 02138
FOREWORD
On May 12, 1978,
the House Select Committee on Assassinations asked Bolt Beranek and Newman Inc.
(BBN) to conduct a preliminary review of the following material:
- Tape recordings
reportedly made of the sounds in Dealey Plaza around 12:30 pm on November 22,
1963
- Transcripts of
the testimony of earwitnesses who were in the Plaza at that same time.
The purpose of this
review was to determine which, if any of this material constituted potential
evidence with respect to the gunfire associated with the assassination of
President John F. Kennedy.
The review
established that (1) only two of the recordings constituted potential evidence
and (2) a statistical analysis of the ear witness testimony could reveal
whether the concept of one rifle is consistent with these individual accounts.
The two tapes found
to be made of the events surrounding the assassination were records from
Channels 1 and 2 of the Dallas Police Department's (DPD) radio dispatching
system.
The Channel 1 tape
contains a continuous record of the sounds transmitted between 12:28 and 12:34
pm over a DPD motorcycle radio stationed in Dealey Plaza. The Channel 2 tape is
an particular, communications between the Chief of the Dallas Police
Department, who occupied the car immediately preceding the Presidential
limousine in the motorcade, and the Channel 2 Dispatcher at DPD headquarters.
An initial analysis
of a portion of the Channel 1 tape did not rule out the possibility that the recording
contained the sounds of gunfire. The House Committee therefore authorized BBN
to conduct studies both of the DPD tapes and of the ear-witness testimony. This
report describes the results of an analysis of the tapes. The study of ear
witness testimony is reported under separate cover. (Green, D.M., "Analysis of Earwitness Reports Relating to
the Assassination of President John F. Kennedy," BBN Rep. 4034, January
1979.)
ACKNOWLEDGMENT
The authors
gratefully acknowledge the fine contributions made to this study by Joseph F.
Colaruotolo, Daniel N. Kalikow, Nancy M. McMahon, Theodore L.
TABLE OF CONTENTS
LIST OF FIGURES
1. INTRODUCTION AND
SUMMARY
The House Select
Committee on Assassinations authorized
Bolt Beranek and
Newman Inc. (BBN) to study two tape recordings made by the Dallas Police
Department (DPD) on November 22, 1963 on Channels 1 and 2 of the DPD's radio
dispatching system. Channel 1 is the channel ordinarily used to handle DPD
radio traffic, and this channel is recorded continuously on a Dictabelt
recorder. Channel 2, an auxiliary channel generally used to handle the
additional radio traffic necessitated by special events, is recorded
intermittently on a Gray Audograph recorder, as actuated by voice
communications and time annotation. Frequent time annotations -- usually at
1-minute intervals are made by the radio dispatchers handling each of these
channels.
On November 22,
1963, during the time of President Kennedy's assassination, the radio of a DPD
motorcycle, which may have been in the motorcade, was stuck in the transmitting
mode on Channel 1 for approximately 5 minutes. During this time, the Chief of
the Dallas Police Department, whose car immediately preceded the President's
limousine in the motorcade, transmitted several messages concerning the
progress of the motorcade over Channel 2. Channel 2 had been designated for use
by DPD officers in the motorcade on November 22, 1963. Therefore, if the
Channel 1 recording were to contain sounds of gunfire associated with the
assassination, then at least one of the motorcycle radios used in the motorcade
must have been incorrectly switched to Channel 1. Voice transmissions on both
tapes were monitored for the call numbers of the 18 motorcycle officers in the
motorcade. Six of the officers were heard to transmit on Channel 2; three on
Channel 1 (These three transmissions were made at about 2:10 pro, 4:39 pm, and
522 pro, all later times than the assassination). The other nine did not make any transmissions, so cannot be
determined which channel their radios were set for.
1.1 Initial
Analysis
The questions to be
addressed in the analysis of these tapes were:
*Does the 5-minute
segment recorded on Channel 1 contain the sound of gunfire?
*If so, how many
shots were recorded and from what location (or locations) did the shots
originate?
To begin with, if
gunfire had been recorded on Channel 1, the analysis of that tape could be
expected to reveal patterns of transient waveforms that would be generally
characteristic of the shock wave produced by the bullet, of the loud and
impulsive noise of the muzzle blast, and of echoes of each. It could further be
expected that the major components of the shock wave would appear in the 1-kHz
to 3.2-kHz frequency band.
The initial
analysis of the Channel 1 tape therefore consisted of filtering and recording
the entire 5-minute segment through each of two filters designed to reveal the
presence of any transient impulsive waveform patterns that might be masked by
the repetitive loud noise of the motorcycle. The first was a bandpass filter
that filtered out all sounds not contained within the frequency range extending
from 1 kHz to 3.2 kHz.
This range was
known to contain the principal frequency components of the shock wave produced
by the bullet and to contain relatively few components of motorcycle noise. The
second filter was an adaptive Widrow LMS filter, which studies the repetitive
nature of noise, estimates what it will be a short time later, and subtracts
these noise components out, leaving transient events not anticipated by the
filter.
The recorded
outputs from both filters for the full 5 minutes were compared, examined, and
plotted on a scale where 5 in. equals 1/10 sec. These plots revealed five
impulse, patterns introduced by a source other than the motorcycle.
Upon closer
examination, all but one of the these patterns were sufficiently similar .to
have had the same source, and the impulses contained in these patterns appeared
to have shapes similar to the expected characteristics of a shock wave and of a
muzzle blast. The remaining pattern was sufficiently different in amplitude and
duration as to have been caused by a different source.
The hypothesis to
be tested, then, was that these four impulse patterns were caused by gunfire.
Initially this hypothesis was subjected to five simple, but necessary,
screening tests:
1. Time of
occurrence
2. Uniqueness of patterns
3. Time span between patterns
4. Shape of impulses within the patterns
5. Amplitude of impulses.
Should the
hypothesis then pass these tests, a sixth, more rigorous, test would be
applied. This final test would require an acoustical reconstruction of the
circumstances of the original gunfire in Dealey Plaza to reveal the relative
times that muzzle blast and shock wave impulse, together with their echoes,
would arrive at microphones located where the motorcycle radio might have been.
1.2 Screening Tests
The five screening
tests were designed to determine whether the characteristics of the four
impulse patterns corresponded both to other evidence and to the characteristics
of actual gunfire.
1 Did the impulse
patterns occur at the same time the shots were actually fired? Yes.
Stopwatch timing
and examination of both tapes placed the time of the shot and the time of onset
of the first pattern of waveforms within 35 sec of each other. The margin of
acceptable time difference was 60 sec, since the two time clocks used by the
two dispatchers were synchronized to within Just 1 minute.
2 Were these
impulse patterns unique? Yes.
Examination of the
entire 5-minute segment did not reveal sufficiently similar impulse patterns
elsewhere on the tape to discount gunfire as the source of these four patterns.
3. Did the time
span between the patterns correspond to other evidence of intervals between
shots? Yes.
The intervals
between the onset times of the four impulse patterns on the DPD tape with the
frames on the Zapruder film showing bullet impact were compared. According to
the Zapruder film, the time span between the earliest and the latest
gunfire-like events recorded on Channel 1 had to be no less than 5.6 sec. The
span between onset times of the first and the fourth patterns was 8.3 sec.
4. Did the shape of
the impulse patterns resemble those generated by actual rifle fire? Yes.
Tape recordings of
test shots made with a Mannlicher-Carcano rifle were put through electrical
circuits that mimicked those through which the 5-minute segment had been
recorded. The shape of the impulse patterns on the Channel l tape approximates
those produced by the test shots.
5. Did the range of
amplitude (loudness) of the impulse patterns resemble that of the echo patterns
produced by the test shots? Yes.
Processing the echo
patterns of the test shots through a radio receiver like that used in the DPD
recording system showed similar compression of the range of amplitude of
recorded signals with respect to the range of the signals fed into the
receiver.
The answers to
these five questions neither proved nor disproved the possibility that the four
impulse patterns on the Channel l tape had been caused by gunfire. A more
rigorous analysis was required to determine with some confidence whether or not
these patterns had been caused by gunfire.
1.3 Further
Analysis
The gunfire and the
potential motorcycle radio positions on November 22, 1962 were acoustically
reconstructed on August 20, 1978 in Dealey Plaza. The sounds were subsequently
processed into echo patterns, each one representing the unique "fingerprint"
of gunfire sounds as heard at one location when a weapon is fired from one
place to one target. The Channel 1 recording made at the time of the
assassination had been similarly processed into sound impulse patterns.
However, the Channel 1 impulse patterns were like badly smudged
"fingerprints," because of the extremely noisy environment in which
the original recording had been made.
The echo patterns
were compared to the impulse patterns to see if any of the clear
"fingerprints" obtained during the reconstruction matched any of the
smudged "fingerprints" on the Channel l recording. The matching
process was a binary correlation detector, a simple but powerful
signal-detection scheme that is conducted mathematically.
Several echo
patterns from the acoustical reconstruction matched sufficiently well with the
four impulse patterns that we were able to place the motorcycle behind the
Presidential limousine, at distances varying from 120 ft to 160 ft.
The correlation
detector indicated that four shots may have been fired, as follows:
l. time 0.0 sec
--one shot from the Texas School Book Depository (TSBD) aimed between the
limousine positions seen in frames 160 and 313 of the Zapruder film
2. time 1.6 sec
-one shot from the TSBD aimed near the limousine position seen in frame 313
3. time 7.8 sec -
one shot from behind the fence on the knoll aimed near the limousine position
seen in frame 313
4. time 8.3 sec
-one shot from the TSBD aimed between the limousine position seen in frame 313
and the triple underpass.
1.4 Conclusions
Based on Results of the Acoustical Reconstruction
The conclusions
drawn from the results of the matches obtained by our analysis were presented
at the public hearing before the committee on September 11, 1978. Essentially,
we had concluded that the motorcycle had indeed been in the motorcade and that
possibly four shots had been fired at President Kennedy. The reason that our
findings with respect to the four shots were stated in terms of probabilities
is as follows.
The correlation
detector produced several false alarms that could be identified as such. These
false alarms are spurious matches caused by uncertainty of the exact motorcycle
position with respect to the known positions of microphones used in the
reconstruction test. Therefore, some of the correlations that indicated the
four shots must also be suspected as false alarms. This uncertainty introduced
by the suspected false alarms can be expressed as a set of probabilities on the
possible true outcomes. These probabilities were calculated from the judgment
that each match has a 50% probability of being a false alarm and from the
assumption that each match is an independent observation. Thus, the individual
probabilities that the shots occurred at each of the four times are:
Shot 1. 88% based
on three matches
Shot 2. 88% based on three matches
Shot 3. 50% based on one match
Shot 4. 75% based on two matches
The probability
that the four possible shots found by the correlation detector include at least
two correct detections is high, about 96%. The probability that there are three
correct detections is lower, about 75%. The probability that all four are
correct is only about 29%. The combined probability that there are three
correct detections, and that the third (knoll) shot is among them is about 47%.
1.5 Independent
Analytical Extension of the Reconstruction Test.
The Committee
sought to have the uncertainty in the test results reduced, particularly with
respect to the 50% probability of the third (knoll) shot. Professor Mark Weiss
and Mr. Ernest Aschkenasy of Queens College were authorized by the Committee to
conduct an analytical extension of our acoustical reconstruction test. They
first identified the objects in Dealey Plaza that caused each echo that appeared
in the echo pattern we had found to indicate the possible third (knoll) shot.
Next, they calculated how this echo pattern would be modified for receivers in
the neighborhood of the microphone from which the echo pattern was obtained.
Finally, they were able to show that 10 echoes of 12 in one of their calculated
echo patterns matched with 10 sound impulses of 14 on the DPD tape recording
each one to an accuracy of +/- l ms. The first of the 10 matching impulses was
found to occur 7.6 sec after the first impulse indicating the first shot.
We examined the
results of this independent study and Judged both the technique and the
parameters they used to be correct in every detail. We further concluded that
the odds were only about 1 in 20 that their very precise match could have been
achieved by chance - i.e., if the 14 sound impulses on the DPD tape were all
noise and did not include echoes from a knoll gunshot. For this reason, We
conclude that there is a 95% probability that there was a gunshot fired from
the knoll at about 7.6 sec after the first one.
1.6 Findings
The results of our
analysis of the tape-recorded evidence, together with the independent analysis
of the echo-pattern match with the third (knoll) shot, permit the following
findings:
1. The recorded
sounds on Channel 1 of the Dallas Police radio dispatch system probably include
the sounds of four gunshots fired in Dealey Plaza at about 12:30 pm on November
22, 1963.
2. The recorded
gunshot sounds were sensed and transmitted by a police radio mounted on a
motorcycle in the motorcade and positioned at distances ranging from 120 ft to
160 ft behind the Presidential limousine.
3. The first
probable shot was fired at about 12:30:47 from the TSBD. The motorcycle
position was then On Houston St. having only about 3 sec earlier slowed in
preparation for the left turn onto Elm St. No shock wave indicating a
supersonic projectile is seen as a precursor to the sounds of the muzzle blast,
and none is expected, owing to the position of motorcycle with respect to the
expected trajectory of the bullet. Therefore, no conclusion can be drawn about
whether this first acoustic disturbance was due to a rifle or to a sound
impulse as loud as the report of a rifle. However, the sound did originate in
the vicinity of the sixth floor of the TSBD.
4. The second
probable shot was fired about 1.6 sec after the first one(,) also from the
TSBD. At this time the motorcycle was just at the corner of Houston and Elm.
Again, no shock wave is seen as a precursor to the sounds of the muzzle, and,
again, none is expected.
5. The third
probable shot was fired about 7.6 sec (This time was obtained from the
independent study of Weiss and Aschkenasy, and it differs by about 0.2 sec From
the time obtained by our correlation detector) after the first one, and it was
fired from behind the fence upon the "grassy knoll". At this time,
the motorcycle was proceeding westward on Elm St. about 80 ft west of the
intersection with Houston St. An apparent shock wave is seen as a precursor to
the sounds of the muzzle blast. In as much as a supersonic projectile would
show such a precursor when the motorcycle is in this position, the third shot
is probably from a rifle.
6. The fourth
probable shot was fired about 5.3 sec. after the TSBD. The motorcycle was on
Elm St. about 90 ft west of the intersection with Houston St. An apparent shock
wave is seen as a precursor to the sounds of the muzzle blast. Since the
trajectory of the bullet would have been over the motorcycle, such a precursor
would be expected for a rifle shot. Therefore, the fourth shot is probably from
a rifle.
7. Additional
police radio transmissions are intermittently probable shots. These
transmissions contribute a few electrical impulse to the noise background in
which the impulses of gunfire are set. However, these noise impulses are too
few in number to have a material effect o n the accuracy by which the echo
patterns of the acoustical reconstruction match the impulse patterns on the DPD
tape.
These findings were
presented at public hearing before the Committee on December 29, 1978. At that
hearing, Officer H.B. McLain of the DPD testified that he had been riding his
motorcycle on the left-hand side of Houston St., approaching Elm St. when he
heard a single shot. After the hearing, he said that he remembered that he had
turned on his siren shortly after the assassination and moved with the
motorcade to the hospital. However, the appearance of McLain in photographs
taken in Dealey Plaza Just after the assassination suggests he did not leave
the area with the motorcade. Unless McLain turned on his own siren, the absence
of the siren sound on the tape is consistent with McLain's behavior as
documented in photographs and it may have been his motorcycle.
Section 2 of this
report describes the acoustical nature of gunfire -- i.e., what could be
expected after appropriate filtering of the Channel 1 tape, if it did indeed
contain the sound of gunfire. Section 3 reports the procedures used to process
the tape and the results of this processing. Section 4 describes the five
screening tests, and Sec. 5 reports the results of the acoustical
reconstruction of gunfire in Dealey Plaza. Section 6 discusses additional
relevant sounds on the Channel 1 recording. Finally, Sec. 7 describes our
review of independent analysis of the match between our acoustical
reconstruction and the sounds of the probable third shot.
2. NATURE OF
RADIO-TRANSMITTED SOUNDS OF GUNFIRE
2.1 Overview
The discharge of a
rifle creates two sources of impulsive sound - the sound of the muzzle blast
and the sound of the shock wave shed from the supersonic bullet as it travels
at a speed greater than the speed of sound. Figure 1 illustrates the difference
in how these two impulsive sounds travel through the air. The shock wave, for example,
has a direct path of travel that resembles a cone, while the sound of the
muzzle blast spreads spherically from the source.
In addition to
traveling at different speeds and in different ways, these impulsive sounds
travel over several different paths before arriving at a receiver -- in this
case, a microphone. Figure 2 illustrates these paths. The first sound impulses
to arrive travel in a straight line from the source to the microphone; this
sound path is called the direct (D) path. It includes reflections (D2) from
impulses traveling the direct path and striking the ground very near the
microphone. Later sound impulses arrive at the microphone after first
reflecting from large surfaces, such as building facades and the ground; these
sound paths are called reflected (R) paths. Even later sound impulses arrive at
the microphone after first diffracting from the corners of buildings and the
edges of other large objects; these sound paths are called diffracted (T, M, L)
paths. A weaker set of sound impulses, arriving at the microphone just after
the direct arrival, are scattered first by small objects such as poles, people,
and automobiles. After striking these scattering objects, these weaker sound
impulses arrive at the microphone over the scattered (S, P) paths. Finally,
reflections from distant objects (U) arrive over various reflected paths, but
these signals appear much later than those arriving by all the previously
described paths.
Fig. 1. LOCI OF
MUZZLE BLAST AND SHOCK WAVES AT TWO TIMES AFTER FIRING OF BULLET.
Fig. 2. ECHO
PATTERNS CAUSED BY DIRECT, REFLECTED, DIFFRACTED, AND SCATTERED IMPULSIVE
SOUNDS IN AN URBAN ENVIRONMENT
All sound impulses
arriving at the microphone that are loud enough to be heard over the
environmental noise would be transmitted over the radio connected to the
microphone. In this case, the environmental noise consisted primarily of the
very loud, repetitive noise made by the engine of a moving motorcycle. This
noise was found to be only about 10 dB lower than the loudest gunfire impulse
"recorded. Thus, only the very loudest gunfire sound impulses would
actually be detectable above the engine noise.
The loudest sound
impulses from gunfire are considerably louder than the loudness of speech, for
which the radio was designed to operate. These loud impulses overdrive the
radio circuitry. Because of the limiting circuits in the radio transmitter,
very loud sounds are recorded in distorted fashion and appear as much weaker
signals than they really are. In fact, despite the difference in loudness of
signals traveling over the several paths illustrated in Fig. 2, each is
recorded as having about the same amplitude.
After the sounds
that were picked up at the microphone had been transmitted to the DPD radio
receiver, the output of the receiver was recorded on a Dictabelt recorder. The
circuitry of the receiver and the characteristics of the recorder also affected
the transmitted signals. The recorded loudness of the sounds transmitted from
the motorcycle radio with the stuck microphone were additionally affected
somewhat by simultaneous transmissions from other officers in the motorcade. An
FM radio receiver, such as the one in DPD headquarters, receives best from the
transmitting radio having the strongest transmitted signal and can accommodate
at the same time all receivers whose transmitted signal strengths differ by
less than the receiver capture ratio.
Thus, the effects
of severe environmental noise, Of the limiting circuitry of the radio
transmitter, of simultaneous radio transmissions, and of the recording
characteristics of a Dictabelt recorder were such that any waveforms that would
emerge from an analysis of the tape would be severely distorted. What these
waveforms would look like without such distortion is illustrated in Fig. 3.
The upper portion
of the waveforms shown in/this figure were produced by a Mannlicher-Carcano
with Western Cartridge Co. ammunition and picked up by a microphone positioned
30 ft from the muzzle and 10 ft to one side of the bullet's trajectory. The
muzzle-blast waveform reveals a peak pressure impulse having a sound pressure
level of 137 dB re 2x10 -5N/m2. For comparison, Fig. 3 also shows the
corresponding waveforms for an M-1 rifle. Despite the differences in loudness
(amplitude) from one weapon to the other, the shock wave and the muzzle blast
can be seen to have characteristic shapes. Sounds processed from the Channel 1
tape could be expected to contain these shapes, but in distorted fashion. The
shapes could be expected to be compressed in amplitude and to be accompanied by
indications of overdriving of the radio circuits. They would also be
accompanied by waveforms produced by the arrival of sound echoes from several
sources, as described in the rest of this section.
FIG. 3. MUZZLE
BLAST AND SHOCK WAVEFORMS FOR MANNLICHER-CARCANO AND M-1 RIFLES.
2.2 Propagation
Over the Direct Path
The distance from
the muzzle in the TSBD to the nearest possible location of the motorcycle
microphone is 60 ft and to the farthest possible location (at Houston and Main)
is 260 ft. Loss in amplitude of the sound of the muzzle blast over the direct
path is due principally to the spherical spreading of the sound as it travels
outward from the source of gunfire. This weakening (attenuation) is accounted
for by the quantity 20 log(D/30), where D is the length, in ft, of the path of
travel. The estimated loudness of the muzzle blast at the nearest possibly
motorcycle location is 137 - 20 10g(60/30), which is equal to 131 dB re 2x10
-5N/m(2). The estimated loudness of the muzzle blast at the farthest possible
location is equal to 118 dB re 2x10 -5N/m(2).
Thus, both the
muzzle blasts and the shock waves would be received over the direct path with
sound pressure levels greater than the approximately 100-dB limiting sound
pressure levels of the motorcycle radio. The result would be both an indication
of overdriving the system and a compression of the recorded amplitude.
2.3 Propagation
Over Reflected Paths
Ground reflections
will always occur from below the microphone at the specular reflection point.
Since the path length of the reflected path is only a few feet longer than for
the direct path, the amplitude of ground-reflected sounds will nearly equal the
amplitude of sounds arriving over the direct path.
Building
reflections occur only when a building facade includes a specular reflection
point for the source and microphone. This condition is met by the buildings on
Houston St. for microphones located on Houston near Main St., and it is also met
by the Post Office Annex for microphones located on Elm St. The path length for
these reflections is the total distance from the source to the specular
reflection point and then to the microphone. For microphones on Elm, the path
length for reflections off the Post Office is about 1100 ft. The amplitude of
such echoes is, therefore, estimated to be 137 - 20 log(1100/30) (= )106 dB re
2x10 -5N/m(2 )-- still loud enough to cause limiting by the radio.
All reflected
sounds, regardless of the reflecting surface, arrive at the microphone T
seconds later than sounds traveling the direct path. T can be expressed as the
ratio AD/C, where AD is the difference between path lengths in ft, and c is the
speed of sound in ft/sec. At 65 deg F, c is 1123 ft/sec, and at 90 deg F, c is
1150 ft/sec. Sounds reflected from the Post Office occur about (1100-100)/1100,
or about 0.9 sec later than the direct sounds.
2.4 Propagation
Over Diffracted Paths
The amplitude of
sound diffracted by a corner of a building can be estimated as follows (See
J.J. Bowman, T.B.A. Senior, P.L.E. Uslenghi, Electromagnetic and Acoustic
Scattering by Simple Shapes, North-Holland Publishing Company, Amsterdam,
1969 p.274). The ratio of diffracted sound pressure Pd to direct sound
pressure Po can be written as: where 6 = r/r0, the distances from the corner to
the source and from the corner to the microphone, respectively. The angle
between arriving and diffracted rays of sound is 0 and k is the acoustic wave
number. The function F is a number generally between 1 and 2.
There are many
corners that can cause diffractions. The corner of the Records Building is
typical. The amplitude of a sound impulse diffracted from its corner and
received at Houston and Elm would be about 30 dB lower than that of an impulse
arriving directly from the source. Since the amplitude of the direct-path sound
of the muzzle blast near Houston and Elm is about 131 dB re 2x10 -5N/m2, the
amplitude of the diffracted impulse will be about 101 dB re 2x10 -5N/m2, still
loud enough to be somewhat limited by the radio and to be quite audible.
The total path
lengths of diffracted sounds vary continuously between limits set by the direct
path length and by the longest reflected path length. Thus, diffracted sounds
should occur between the time of the direct arrival and the time of the arrival
of the reflection from the Post Office.
2.5 Propagation
Over Scattered Paths
Objects small
enough so that kd=2, where d is the nominal diameter of the object, will
scatter sound in all directions. Substantial energy in the muzzle blast impulse
is contained at frequencies near 500 Hz, where k = 2.8 ft -1. Thus, objects
having a diameter of about l ft satisfy the scattering requirement. Such
objects could be light poles, people, and motorcycles.
The loudness of
scattered sound diminishes rapidly with increased distance from the scattering
object. For this reason, only sounds scattered from objects fairly close to the
microphone would be loud enough to be recorded.
Scattered sounds
loud enough to be picked up by the microphone would arrive just following
strong direct, reflected, and diffracted sounds. These scattered arrivals tend
to increase the apparent time interval in which the primary signals arrive.
3. RESULTS OF
EXAMINING AND PROCESSING THE DPD CHANNEL 1 TAPE
The first tape we
received on May 12 from the Committee had a very scratchy overlay of needle
noise, indicating that it was a very poor or multiple-generation dub of a
recording. In july, the Committee gave us an electromagnetic tape recording
that was identified as an original dub made by the DPD, as well as the original
Dictabelt record. We then made our own dub on magnetic tape from the original
Dictabelt record and compared our dub with that reportedly made by the DPD. We
digitized both dubbed tapes - ours and that made by the DPD, plotted the
outputs of the digitizing process, and found them to be virtually identical. In
this way, we determined that the Dictabelt record was really the source of the
data on the DPD-dubbed tape that we were using for analysis.
On the DPD Channel
1 tape, there is an interval of about 5-minute duration, beginning a little
after 12:28 pm, in which the radio traffic on this channel is disrupted by a
continuous transmission by some remote transmitter, presumably because its
transmit button was stuck in the "on" position. As described in
Appendix A, we input this entire interval into a digital computer, for
subsequent detailed listening, viewing, and processing. This section describes
the results of that examination.
3.1 The Unprocessed
Waveform Data
First, we made a
high-resolution graphical plot of the waveform of this signal, at a scale of 5
in. per 1/10 sec, for detailed visual examination. The plot of the entire
interval comprises a roll of paper 12 in. wide by 234 ft long. Reductions of
excerpts of this plot are reproduced in Fig. 4. In this figure and in the
following discussion, time is noted in seconds from the beginning of the
interval.
The first region to
be noted in Fig. 4 is the area around 131 sec. This region is typical of the
high level of motorcycle noise that characterizes the first 2 minutes of the
data.
In the region of
132 to 133 sec, we can see the amplitude of the noise slowly drop. Later, when
we discover the trajectory of the motorcycle as a by-product of detecting the
sounds of shots, we find that the motorcycle was approaching the corner of
Houston and Elm Sts. at this time. Therefore, this diminution of motorcycle
noise is probably due to the slowing necessary to negotiate the 120(deg) left
turn at the corner.
At about 36.5 sec,
we note a single large impulse of relatively long duration. Because of its
length and because the region following this impulse is largely free of other
impulses, such as the echoes normally associated with loud impulsive sounds, we
feel that it is unlikely that this impulse represents the sound of gunfire.
The regions around
137.3 to 138.7 and 139.2 to 140.9 sec are notable for a number of brief, loud
impulses. These impulse patterns, the first to appear in the data up to this
time, were judged as potentially representing gunfire.
The region from
144.8 to 147.2 sec, which does not appear in Fig. 4, also contains a large
number of impulses of similar character. Because this region is about twice as
long as the preceding ones. it was identified as possibly representing two
separate impulse patterns, and, therefore, as potentially containing the sounds
of two shots.
FIG. 4. WAVEFORMS
RECORDED FROM CHANNEL 1 TRANSMITTER WITH STUCK MICROPHONE.
3.2 Spectrographic
Analysis
Another way of
portraying acoustical data is in the form of a spectrogram, in which the
short-term spectrum of the signal is displayed as a function of time. Two
example spectrograms from the region 141 to 148 sec are shown in Fig. 5. In
this figure, time runs from left to right across the figure, and frequency from
bottom to top. The energy at a given time and frequency is depicted by the
blackness of the paper at that point.
The region from 141
to 144 sec is only noise. Just after 144 see, a single loud click occurs,
followed by a region of very faint speech (faint diagonal and horizontal
smudges that change rapidly), clicks (thin vertical lines), and keying
heterodynes (steady horizontal bars). The analysis into characteristic
frequency components performed by the spectrograph permits us to recognize
these events in a way not possible in the waveform patterns.
3.3 The Filtered
Waveform Data
To be sure that the
137- to 147-sec region of the transmission contained the only transients of
potential importance with respect to gunfire, we attempted to remove the effect
of the motorcycle engine noise to see if it was obscuring other transients. For
this purpose, we implemented on a high-speed digital computer a noise-canceling
filter program that adapts to and subsequently cancels sound components that
appear to be non-random (in this case, the periodic noise of the engine). This
filtering algorithm is described in Appendix A. It was tested on a
high-fidelity recording of motorcycle engine noise and was found to be very
effective in removing it.
FIG. 5. SPECTROGRAMS
FROM WAVEFORM RECORDED FROM CHANNEL 1 TRANSMITTER WITH STICK MICROPHONE.
The adaptive
filtering algorithm, when applied to the entire 5-minute segment of
transmission, was not so effective. Figures 6 and 7 show the effect of
filtering the waveform from 130 to 150 sec (overlapping the period for which
the unprocessed waveform is shown in Fig. 4). The adaptive filtering removed
hum and some low-frequency noise components, but the overall effect was not
dramatic. Evidently, the distortions introduced by the radio transmitter, the
original Dictabelt recording system, and the subsequent multiple playings of
the Dictabelt had added nonrandom noise components that the adaptive filter was
unable to remove.
Appendix A also
describes other signal-processing techniques that were applied to these data in
attempts to remove the motorcycle noise and to detect and track motorcycle
engine speed. The results in both cases were negative.
FIG. 6. ADAPTIVE
FILTERED WAVEFORMS RECORDED FROM CHANNEL 1 TRANSMITTER WITH STUCK MICROPHONE
(130 to 141 sec.)
FIG. 7. ADAPTIVE
FILTERED WAVEFORMS RECORDED FROM CHANNEL 1 TRANSMITTER WITH STUCK MICROPHONE
(141 to 150 sec).
4. SCREENING TESTS
As described in
Sec. 1, the four impulse patterns on the DPD tape were subjected to five simple
but necessary screening tests. If the patterns did not pass any of these simple
tests, then they could safely be assumed to have been caused by something Other
than gunfire. If they were to pass these tests, they could not be assumed to be
gunfire, but further analysis would be warranted. Essentially, the screening tests
were designed to answer the following questions:
l. Did the impulse
patterns occur at the same time as the assassination?
2. Were the
patterns unique? In other words, were they caused by the same source, and did
they appear only at this time and nowhere else on the tape?
3. Did the time
intervals between the impulse patterns match that of other evidence of gunfire?
4. Did the shape of
the impulses resemble the shape of impulses of recorded gunfire?
5. Was the
amplitude of the impulses similar to that of recorded gunfire?
This section of the
report describes how these questions were answered.
4.1 Time of
Occurrence
To determine the
time Of day when the impulse patterns were recorded on Channel l, we examined
the Channel l and the Channel 2 tapes. It is usual DPD practice for the
Dispatchers on both channels to make frequent time annotations. In doing so,
they refer to two different clocks, which are synchronized at the beginning of
each month and which are read out in full minutes only. An FBI study concluded
that, towards the end of the month, the clocks could differ by as much as 1
minute. The allowable difference in the timing of events on Channels l and 2,
therefore, was 60 sec.
The Channel 1
segment was a continuous recording that had no time annotations during the
period of stuck transmission, but time annotations preceded and followed this
period. The Channel 2 segment was an intermittent recording with frequent time
annotations throughout. A stopwatch was used to time the events on both channels.
Figure 8
illustrates the results of stopwatch timing of the Gray Audograph record of
Channel 2 events. Time annotations made by the Channel 2 Dispatcher are plotted
against time on the stopwatch for the interval extending from 12:22 pm to 12:40
pm. Lines representing the least-square error fit are drawn through the time
annotations. Note that the clock used by the Dispatcher is read out only in
full minutes, and occasionally there is more than one annotation for the same
minute.
For the events occurring
before 12:30 pm on the Channel 2 tape, the slope of the least-square error fit
is only 0.4, indicating intermittent operation of the recorder, which stops
recording when there are no Voice transmissions. At about 12:30 pm, the voice
traffic picked up, and the Gray Audograph began recording continuously, as
indicated by a least-square error fit slope of 1.0.
FIG. 8. LEAST
-SQUARE ERROR FITS TO CHANNEL 2 DISPATCHER'S TIME ANNOTATIONS SHOWING TIMES OF
DPD CHIEF'S RADIO TRANSMISSIONS.
The stopwatch time
of two successive transmissions from Chief Curry are noted at the left of the
illustration between the period extending from 6 minutes to 8 minutes. In the
first, he notes that the motorcade is "approaching the triple
underpass." After the Dispatcher notes the time as being 12:30, the Chief
announces, "We are going to the hospital, officers." The
assassination must have occurred sometime between Chief Curry's two voice
transmissions. Since the slope of the least-square error fit changes at about
12:30, it is impossible to determine precisely the time on the Channel 2 clock
when the assassination occurred. The best estimate is 12:30:12 pm.
Figure 9
illustrates the results of stopwatch timing of the Dictabelt record of the
events on Channel 1. Here, the slope of the least-square error fit is 0.95,
indicating that the recorder was running 5% too slow and, therefore, was
compressing time slightly. (Frequency
analysis of the power hum on the tape recording also indicated that the
recorder had been about 5% slow. Since the hum could have been added when the
tape was recorded from the dictabelt, this is not a reliable indication of the
original recording speed.) The fact
that the slope does not change over the course of the entire segment shows that
the recorder operated continuously.
The onset of the first impulse pattern, or gunfire-like event, on Channel 1 occurred at 12:30:47, Channel l time. Thus, the events on Channels l and 2 occurred within 35 sec of each other, well within the time difference allowable for this screening test.
FIG. 9. LEAST -
SQUARE ERROR FIT TO CHANNEL 1 DISPATCHER'S TIME ANNOTATIONS SHOWING TIME OF
FIRST SET OF GUNFIRE - LIKE EVENTS.
4.2 Uniqueness of
the Impulse Patterns
If impulse patterns
similar to those occurring at the time of the assassination were to be found
anywhere else during the 5-minute recording of stuck transmission, then the
patterns could safely be assumed to have been caused by something other than
gunfire. Thus, we examined processed waveforms for the entire segment of stuck
transmission, looking for impulse patterns similar to those already identified.
During the course of this examination, only one other pattern was found. It
began about 30 sec after the other four patterns and was comprised mostly of
impulses apparently caused by radios keying in, attempting to transmit. This
sequence, which lasted for approximately 4 sec, did not resemble the earlier
impulse patterns well enough to have been caused by the same source.
4.3 Time Span of
the Impulse Patterns
If the impulse
patterns were caused by the gunfire of the assassination, the time span they
occupy would have to be at least as long as the evidence of time between bullet
impacts as seen on the Zapruder film. On that film, bullet impact is Judged to
occur before frame 210 and again at frame 313, an interval of 103 frames. Since
Zapruder's camera was judged to be operating at 18.3 frames per sec, the time
span between these two events is 5.6 sec. The time span between the onset of
the first impulse pattern and the onset of the fourth impulse pattern on the
Channel 1 tape is 7.9 sec. When corrected for the fact that the tape recorder
was running about 5% too slowly, the real time span is 8.3 sec,
FIG. 10.
MUZZLE-BLAST AND SHOCK WAVEFORMS TRANSMITTED BY A POLICE RADIO SIMILAR TO THE
ONE USED BY DPD MOTORCYCLES FOR SEVERAL DIFFERENT LOUDNESSES.
4.4 Shape of
Impulses
If the impulse
patterns recorded on the DPD tape were gunfire, the shape of the waveforms
would have been distorted by the limiting circuitry of the radio transmitter.
Figure 10 shows the nature of these distortions. At the left of the figure is a
muzzle-blast waveform obtained from the test firing of a Mannlicher-Carcano
rifle. This waveform has a double peak showing the direct arrival of the muzzle
blast with a strong ground reflection immediately following. A tape recording
of these impulses was fed through a transmitting and recording system similar
to that used by the DPD. The characteristics of both these systems are
discussed in Appendix B.
The series of five
photographs of transmitted muzzle-blast waveforms shows the effect of the
system's circuitry on impulse shapes --essentially, the louder the input
signal, the greater the distortion. For example, the top photograph shows how
the loudest signals, those arriving over the direct path, would be recorded.
The signal that was input at 109 dB is a good example of what the reflection
from a large and distant surface, such as the Post Office, would look like.
Similar analysis of the shock-wave impulse at the right of the figure
illustrates how the simple N-wave of the bullet is severely distorted when the
input signal greatly exceeds the 100-dB limiting circuitry of the transmitter.
Comparison of these
waveforms with the impulse patterns obtained from the DPD tape showed
sufficient similarity that the possibility that the impulse patterns were
caused by gunfire could not be ruled out.
4.5 Amplitude of
Impulses
Another
characteristic of the waveforms that would have been affected by the circuitry
of the radio transmitter if the input signal was as loud as gunfire was their
amplitude. The recorded amplitudes of the sounds would be compressed in such a
way that strong signals would appear to be weaker than they actually were, and
weak signals in the same pattern would, therefore, appear stronger. As can be
seen in Fig. 11, this compression is greatest for very loud signals, especially
those with high-frequency content. For example, although all the signals were
compressed, the amplitude compression of muzzle-blast waveforms above 100 dB
was in every case less than that of the shock waves that are of higher
frequency.
When the
peak-to-peak difference in amplitude between two signals was 30 dB, they were
recorded as having only a 20-dB difference (muzzle blast) or only a 10-dB
difference (shock wave). As the amplitude of the input signal decreased, the
difference in peak-to-peak level became more noticeable. This analysis gave us
greater insight into the characteristics of the sounds originally recorded on
the DPD tape. The signals on that tape also appeared to be compressed in
amplitude, indicating that the sounds, as originally picked up at the
motorcycle microphone, may have been loud enough to have been caused by
gunfire.
FIG. 11. LEVEL OF
TRANSMITTED WAVEFORMS AS A FUNCTION OF WAVEFORM LEVEL AT THE MICROPHONE.
5. ACOUSTICAL
RECONSTRUCTION IN DEALEY PLAZA
Because the five
screening tests described in Sec. 4 had failed to disprove the possibility of
gunfire having been recorded on the Channel 1 tape, a more rigorous test was
required. The objective of the acoustical reconstruction, therefore, was to
obtain several "acoustical fingerprints" of the sound of gunfire in
Dealey Plaza to compare with the impulse patterns found on the Channel l tape.
If any of the "fingerprints" matched, then the reconstruction would
result in determining both the timing of the shots and the locations of the
weapon and the target for each shot. Only those weapon and target locations
indicated by available testimony were to be tested.
5.1 Nature of the
Test
The most powerful
test for the presence of weak signals that have many known features, but that are
not clearly detectable because of background noise, is the correlation
detection test. There are six distinct steps required to conduct this test.
Step 1: Obtain
acoustical measurements called test patterns, of the signals to be detected.
These test patterns
are uniquely determined by weapon- target-microphone locations for each shot.
There were 12 combinations of weapon-target locations, and they are listed in
Table T. There were 36 microphone locations (3 arrays of 12 microphones),
which, along with the four target locations, are illustrated in Fig. 12. Thus,
432 (12 x 36) unique test patterns were obtained. Six of these are illustrated
in Fig. 13, where the logarithm of sound-pressure amplitude is displayed as a
function of time, on a scale 16 in. to 1 sec.
FIG. 12. MICROPHONE
LOCATIONS AT DEALEY PLAZA.
FIG. 13. COMPARISON
OF TEST ECHO PATTERNS PRODUCED BY BOTH WESTERN AND NORMA AMMUNITION FIRED FROM
TSBD (MUZZLE WITHDRAWN) AT TARGET NO. 3 AND RECEIVED AT ARRAY 3, MICROPHONES 7,
8, AND 9.
TABLE I. SEQUENCE
OF TEST SHOTS
Weapon Location |
Target 1 |
Target 2 |
Target 3 |
Target 4 |
TSBD (Muzzle in
plane of window) |
Shot 1 |
Shot 3 |
Shot 6 |
Shot 10 |
TSBD (Muzzle 2 ft
inside plane of window) |
Shot 2 |
Shot 4 |
Shot 7 |
Shot 11 |
Knoll (Rifle) |
Shot 5 |
Shot 8 |
Shot 12 |
|
Knoll (Pistol) |
Shot 9 |
|
|
|
Step 2: Process the
432 unique test patterns into a like number of unique echo patterns or
"fingerprints."
Since the radio
receiver compresses the amplitude of loud gunfire sounds into a narrow range of
amplitudes, for comparison with the compressed impulse patterns, test-shot
echoes that differ greatly in loudness must be compressed so as to differ only
slightly in loudness after transmission by the radio. To achieve this
compression, we selected only those echoes in a pattern having sufficient
loudness to render them distinct from their neighboring weaker echoes.
This echo selection
process is illustrated in Figs. 14 through 17, for test patterns of individual
shots as recorded by three adjacent microphones. For each of these figures, the
geometry of the test shot -- i.e., the weapon-target microphone location
sequence -- can be reconstructed by referring to Fig. 12. As can be seen from
that figure, 12 microphones were placed in 3 successive arrays along the route
of the motorcade, beginning at the right of the figure at the corner of Houston
and Main. The outputs of the microphones were recorded on channels having the
same numbers as the microphones. Thus, the echo patterns in Fig. 14 represent
the sound of gunfire made by a MannlicherCarcano rifle, withdrawn 2 ft within
the plane of the TSBD window, fired at the target located closest to the TSBD,
and picked up by microphones 4, 5, and 6 located on Houston St. before the turn
onto Elm.
In each of these
four figures, 14 through 17, very loud echoes were selected from the echo
patterns recorded by the three adjacent microphones. Those echoes Judged to
have been caused by some feature in Dealey Plaza -- e.g., direct arrivals of
shock wave and muzzle blast, ground and building reflections, etc. were
identified by dots that are connected by nearly vertical lines. The reason the
lines are not vertical is that the microphones were far enough apart to receive
the same sound at dif ferent times. When Fig. 14 is again used as an example,
the slope of the vertical lines at the left of the figure indicates that
microphone 6 was closest to the weapon location and was, therefore, the first
microphone to pick up sound arriving by the direct path and by other short
paths. The slope of the lines at the right of the figure indicates that
microphone 4 was closest to a major reflecting surface, such as the Post
Office, and was the first to pick up those echoes.
From the four
groups of echo patterns shown in Figs. 14 through 17, we selected as
"fingerprint" material the following number of echoes: 15, 14, 9, and
10. Again, selection of these echoes was based on their strength and on an
understanding of how all the echoes would be compressed in amplitude by the
limiting circuitry of the DPD dispatching system. The same procedure was used
to select echo patterns from each of the 432 test patterns. Each echo pattern
consisted of dots placed at the time of each echo on a scale of 16 in. per sec,
and there was an average number of 12 echoes on each echo pattern. Most echo
patterns were of about l-sec duration, or 16 in. long.
FIG. 14. ECHO
PATTERN FOR SHOT 2 (TSBD, MUZZEL WITHDRAWN, TARGET NO. 1) RECEIVED AT ARRAY 2,
MICROPHONES 4, 5, AND 6.
FIG. 15. ECHO
PATTERN FOR SHOT 7 (TSBD, MUZZLE WITHDRAWN, TARGET NO. 3)RECEIVED AT ARRAY 2,
MICROPHONES 4, 5, AND 6.
FIG. 16. ECHO
PATTERN FOR SHOT 8 (KNOLL, TARGET NO. 3) RECEIVED AT ARRAY 3, MICROPHONES 4, 5,
AND 6.
FIG. 17. ECHO
PATTERN FOR SHOT 6 (TSBD, MUZZLE EXPOSED, TARGET NO. 3)RECEIVED AT ARRAY 3,
MICROPHONES 4, 5, AND 6.
Step 3: Process
into impulse patterns the segment of the tape recording that passed all five
screening tests.
The amplitude of
the sounds on each segment was displayed in dB as a function of time, with each
second of data occupying 16 in. of the display. The tape segment was subdivided
for convenience into six separate segments of about 1-sec duration, each
segment containing numerous sound impulses.
(These segments included the four impulse patterns that passed the
screening tests, with the fourth pattern divided into two segments, and one
pattern that did not pass the tests.)
About 4 sec of data were discarded, because there were no impulses
occurring within them. All impulses louder than a threshold value were selected
as members of the impulse pattern. This process is illustrated in Fig. 18,
where 17 impulses were selected in a 1.2-sec-long segment of the DPD tape that
begins at 137 sec from onset of the stuck microphone.
Above each numbered
impulse in Fig. 18 is a pair of vertical lines separated from the time of
impulse by 6 msec. The 12 msec between this pair of lines represents a window
in which an echo from an echo pattern recorded during the reconstruction might
acceptably occur. The reason for establishing such an acceptance window for the
comparison between impulse and echo patterns is that the precise motorcycle
position and, therefore, its position relative to the actual test microphone
locations, was not known. This subject is addressed further in Sec. 5.2.
Three other impulse
patterns are illustrated in Figs. 19 through 21. These correspond to DPD tape
segments that begin at 139, 145, and 145.5 sec, and they contain 15, 11, and 8
impulses, respectively. The two impulse patterns not illustrated contained 4
and 8 impulses, so that all six of the l-sec segments averaged 10.5 impulses
each.
FIG. 18. IMPULSE
PATTERN FROM STUCK-TRANSMITTER RECORDING BEGINNING AT TIME 137 SEC.
FIG. 19. IMPULSE
PATTERN FROM STUCK-TRANSMITTER RECORDING BEGINNING AT TIME 139 SEC.
FIG. 20. IMPULSE
PATTERN FROM STUCK-TRANSMITTER RECORDING BEGINNING AT TIME 145 SEC.
FIG. 21. IMPULSE
PATTERN FROM STUCK-TRANSMITTER RECORDING BEGINNING AT TIME 145.5 SEC.
Step 4: Correlate
each of the 432 echo patterns with each of the six impulse patterns for a total
of 2592 separate correlation coefficients.
The process of
correlation, which obtains the measure of goodness of match between an echo
pattern and an impulse pattern, is the essence of the correlation detector. The
process is carried out by sliding the impulse pattern along the echo pattern
until the maximum number of echoes occurs within the acceptable windows of
corresponding impulses. This maximum number is called the number of matches.
The correlation coefficient is the number of matches divided by the square root
of the product of the number of echoes and the number of impulses; i.e., correlation
coefficient
If there is an
equal number of echoes and impulses, and if they all match when the two
patterns are positioned at one relative time, then the match is perfect and the
value of the correlation coefficient is unity (1.0). If there are extraneous
impulses or echoes, such as may be caused by noise on the DPD tape or by an
echo-acceptance threshold too low for the reconstructed sounds, then the match
cannot be perfect and the correlation coefficient will be less than unity. If
the echo pattern is not at all similar to the impulse pattern, there will be
only one or two matches. and the correlation coefficient will be only a little
larger than zero.
The correlation
coefficients for all 2592 matches were calculated by determining the maximum
correlation coefficient possible for each, after sliding each pair of echo
patterns and impulse patterns relative to one another. The time of the first
impulse on the impulse pattern was noted with respect to the instant that the
microphone button became stuck.
Step 5: Select all
correlation coefficients having values greater than the detection threshold
value.
The detection
threshold concept is necessary because we have observed that noise and
experimental uncertainty tend to prevent any perfect correlations (unity value
of the correlation coefficient). Whenever an echo pattern matches sufficiently
well with an impulse pattern to produce a correlation coefficient higher than
the threshold value, that echo pattern is said to pass the detection test. There
are two possible meanings to be assigned to each passing of the test. First, if
the impulse pattern was truly caused by gunfire, the passing is called a
detection. Second, if the impulse pattern was not caused by gunfire -- but
rather by any other event capable of introducing noise in the radio -- then the
passing is called a false alarm.
Since impulse
patterns that are truly caused by gunfire and mixed with radio noise cannot be
expected to correlate perfectly with the test patterns, the detection threshold
must be set low enough to ensure that no detections are missed. But the lower
the detection threshold is set, the more false alarms that can be expected to
occur. The analysis described in Appendix C indicated that random noise on one
l-sec segment of the DPD tape having about 12 impulses per sec will cause fewer
than 3.3 false alarms out of 432 echo patterns, provided the detection
threshold is 0.6. This rate of false alarms was Judged acceptable and was
considered necessary to prevent misses.
Step 6: Eliminate
from the set of detections and false alarms the false alarms that can be
identified.
Since false alarms
are caused by noise (unexplainable events), they may occur with echo patterns
that represent weapon, target, and microphone positions that are obviously
disjoint from actual detections and from false alarms that mimic actual
detections. These events can be correctly identified as false alarms and
eliminated from further consideration. Only independent (non-DPD tape) evidence
can be used to identify those false alarms that may be mimicking detections.
5.2 Problems To Be
Solved by the Acoustical Reconstruction Test
The acoustical
reconstruction test had to be conducted in a safe and timely manner. Therefore,
all conceivable weapon, target, and microphone locations could not be
reconstructed. Five problems arising from this fact, and their solutions, are
discussed here.
1. Where in the
motorcade was the motorcycle during the time span of the assassination,
assuming that it was in the motorcade?
The motorcycle
would need to be within the confines of Dealey Plaza in order to pick up the
sound of gunfire, From the corner of Main St. and Houston St. to the position
of the President's limousine at the time the President's head wound was
inflicted is a distance of about 460 ft. Since the pavement widens greatly at
the corner of Houston and Elm, about 150 ft needed to be covered twice, for a
total linear distance of 610 ft. It was judged that there would be time from
sunrise until noon to conduct four complete firing sequences. One of these
would need to be a repeat to test for the similarity of two types of
ammunition. Only 12 microphones could be used simultaneously, because of the
need to keep 2 channels of a 14-channel tape recorder in reserve for annotation.
Therefore, 36 microphone positions would have to cover a distance of 610 ft.
Also, the streets in Dealey Plaza are about 40 ft wide, meaning that a
motorcycle would likely be no more than about 18 ft from the center of a
street. For these reasons, the microphones were spaced 18 ft apart, as
illustrated in Fig. 12.
Because of the
spacing of the microphones and lack of knowledge of the precise position of the
motorcycle within the motorcade, it was Judged that the motorcycle would, in
the worst case, have been no more than 18 ft away from a microphone location.
The most likely separations were accounted for in Sec. 5.1, Step 3, by the
establishing of a -+6-msec acceptance window for matching echo and impulse
patterns.
2. Is live
ammunition necessary in the acoustical reconstruction, and does the type of
weapon and ammunition make a difference?
In Sec. 2, we
described how the shock waves generated by rifle bullets would be sufficiently
loud at some microphone positions to become a significant part of an echo
pattern. The speed of the bullet is important, because it determines the
difference in time between perception of the shock waves and perception of the
muzzle-blast waves. Therefore, it would be best to use the same type of rifle
and the same type of ammunition in the reconstruction as was used in the
assassination. Evidence dictated use of a Mannlicher-Carcano rifle from the
sixth floor of the TSBD, firing Western Cartridge Co. ammunition. The Committee
supplied a similar rifle, but only 15 WCC rounds. It was necessary to use Norma
ammunition for the first three sequences of rifle fire, while the fourth
sequence was a duplicate of the third, with the exception of substituting WCC
rounds for Norma rounds. No significant difference due to the type of
ammunition was observed, as can be seen in Fig. 13.
Gunfire from behind
the fence on the knoll had been alleged by some, although there was no evidence
to indicate what type of weapon or ammunition might have been used. The
greatest difference between echo patterns caused by two different weapons
occurs whenever one fires a supersonic projectile and the other a subsonic one.
For this reason, a Mannlicher-Carcano was used to produce the supersonic
projectile and a 38-caliber pistol was used to produce the subsonic projectile.
Since the knoll-to-target distances were only about 100 ft, it was not
necessary to obtain great accuracy in matching test weapons with alleged
assassination weapons in this case.
3. Where should the
targets be located?
Photographic
evidence indicated that shots struck the President when his limousine was at
the locations indicated approximately in Zapruder frames 200 and 313. Also,
evidence indicated that a bullet may have struck the curb on the south side of
Main St., near the triple underpass. Finally, our initial investigation of the
tape indicated a shot may have been fired when the limousine was near frame
160. For these reasons, four targets were selected, and their positions are
marked on Fig. 12.
4. Had any
significant physical changes been made in Dealey Plaza?
An important factor
to be considered was the change over 15 years in the physical, and therefore
acoustical, characteristics of Dealey Plaza. The absence of the crowd and cars
was Judged insignificant, because reflections of sound from these sources would
constitute sufficiently weaker signals than those that would be selected for
analysis. Thus, only two changes of importance had taken place: the
introduction of the Hyatt Regency building a couple of blocks away as a
possible reflector of sound and the absence of the building formerly located at
the southeast corner of Commerce and Houston Sts.
Travel time for a
sound wave to reach the Hyatt Regency and be reflected back to the microphone
was estimated at 2 sec. Since the four impulse patterns had durations of no
more than 1.1 sec each, the echo from the Hyatt Regency would not distort the
data.
The sound waves
that originally hit the missing building would have been diffracted by the
corner of the building itself, with much of the sound energy being scattered.
The reflected signal from this building would, therefore, have been
sufficiently weak to have been swamped by the very strong reflection coming off
the Post Office Building located at the southwest corner of the same streets.
For the purposes of
reconstruction, therefore, Dealey Plaza was judged to have the same acoustical
characteristics in 1978 that it had in 1963. However, when the test was being
set we found that to shoot from the TSBD at Target No. 2, it was necessary to
shoot between two overhead signboards on a sign above Elm St. that was not
there during the assassination. This sign could not be moved. The secondary
echoes generated by the projectile shock waves impinging on these two signs
apparently reduced the correlation coefficients for matches with test shots at
this target, for only three were ever found to exceed the detection threshold,
and these were identified as false alarms.
5. How could the
listening tests be accommodated?
The experts used by
Dr. Green (See footnote, p. iv.) to determine how earwitness accounts of the
sounds of gunfire might be explained needed to hear each of the various test
shots from at least four different locations. This requirement was met by our
use of four identical sequences of test shots.
5.3 Results of the
Acoustical Reconstruction Test
Of the 2592 maximum
correlation coefficients determined by correlating the 432 echo patterns with
the impulse patterns on six tape segments, 15 correlation coefficients exceeded
the detection threshold value of 0.6. The time and weapon-target microphone
locations for each of these coefficients are listed in Table II. Inspection of
the table shows that no correlations exceeded the threshold value for the two
segments beginning at 136.20 sec and 146.30 sec after the time the microphone
button became stuck. Fourteen of the 15 correlations that did exceed the
threshold value occurred at four different instants of time, those beginning at
137.70 sec, 139.27 sec, 145.15 sec, and 145.61 sec. This result shows the
possibility of four shots having been fired, each at One of the four times
listed. The fifteenth correlation value to exceed the detection threshold
occurred at 140.32 sec after the time the microphone button became stuck. This lone
correlation will be identified as a false alarm in the next section and,
therefore, does not indicate the possibility of a fifth shot. These times are
all about too small, because the tape-recording process was found to be about
5% slow (see Sec. 4.1).
List of All 15
Correlations Between Impulse Patterns Occurring in 6 Segments of the DPD Record
and Echo Patterns from 432 Test Shots (2592 Separate Correlations) Having a
Correlation Coefficient Higher than 0.5
Beginning
Time of |
Microphone
Array |
Rifle
|
Target |
Correlation |
136.20
sec |
No
Correlations |
Higher
Than |
... |
0.5 |
|
|
|
|
|
137.70
sec |
2(5) |
TSBD* |
1 |
0.8 |
" |
2(5) |
TSBD* |
3 |
0.7 |
" |
2(6) |
TSBD |
3 |
0.8 |
" |
2(6) |
KNOLL |
4 |
0.7 |
|
|
|
|
|
139.27
sec |
2(6) |
TSBD* |
3 |
0.8 |
" |
2(6) |
TSBD |
3 |
0.6 |
" |
2(10) |
TSBD |
3 |
0.6 |
140.32
sec |
2(11) |
TSBD* |
3 |
0.6 |
139.27
sec |
3(5) |
KNOLL |
3 |
0.6 |
|
|
|
|
|
145.15
sec |
3(4) |
KNOLL |
3 |
0.8 |
" |
3(7) |
TSBD* |
2 |
0.7 |
" |
3(8) |
TSBD |
3 |
0.7 |
|
|
|
|
|
145.61
sec |
3(5) |
TSBD |
3 |
0.8 |
" |
3(6) |
TSBD |
4 |
0.8 |
" |
3(8) |
TSBD* |
2 |
0.7 |
|
|
|
|
|
146.30
sec |
No
Correlations |
Higher
Than |
... |
0.5 |
TABLE II. LIST OF ALL 15 CORRELATIONS BETWEEN IMPULSE PATTERNS OCCURRING IN SIX SEGMENTS OF THE DPD RECORD AND ECHO PATTERNS FROM 432 TEST SHOTS (2592 SEPARATE CORRELATIONS) HAVING A CORRELATION COEFFICIENT HIGHER THAN 0.5.
5.4 Conclusions
about the Acoustical Reconstruction Test
It becomes clear
upon examination of the weapon, target, and microphone locations for the
several echo patterns that passed the correlation detection test at each of the
four different times, that some are inconsistent with each other. Thus, some or
perhaps all represent false alarms. Deciding which are false alarms was greatly
facilitated by plotting the microphone locations for each of the 15 echo
patterns against the time on the DPD tape when it correlated highly. This plot
appears in Fig. 22, where zero on the time scale is taken to be the time on the
DPD tape where high correlations were first detected. Zero on the distance
scale is taken at the point where the Hughes film* shows a motorcycle to be,
just as the Presidential limousine is seen to disappear around the corner from
Houston St. onto Elm St. This motorcycle position is marked M in Fig. 12.
Distance is measured in feet from this point along the motorcade route.
*Frames from the
film taken by Robert Hughes, an amateur photographer, were introduced as
evidence at the December 29 Hearing. This film was taken from the left-hand
edge of Houston St., near Main St. With
the camera pointed north up Houston St., the limousine is seen just
disappearing around the corner after a left turn onto Elm St. A few frames
later a motorcycle passes through the field of view, moving from right to left,
proceeding north on Houston St.
FIG. 22. MICROPHONE
POSITIONS ALONG MOTORCYCLE ROUTE WHERE HIGH CORRELATIONS WERE OBTAINED, AS A
FUNCTION OF TIME. ESTIMATED TRAJECTORIES OF MOTORCYCLE AND OF THE PRESIDENTIAL
LIMOUSINE ARE SHOWN FROM THEIR POSITIONS INDICATED BY THE HUGHES FILM AT THE
TIME THE LIMOUSINE TURNED DOWN ELM ST.
Even a brief glance
at Fig. 22 shows that the microphone locations that correspond to correlations
at the three times after the first impulse tend to progress uniformly forward
along the motorcade route. This conclusion can be quantified statistically by
the chi-square test. If the motorcycle were not moving through Dealey Plaza at
the time of the assassination, the distance along the motorcade route would be
a meaningless coordinate, and the microphone locations for the correlations
that exceed the detection threshold would occur at random. When the chart in
Fig. 22 is partitioned into a 2 x 2 table by separating time at 5 sec and
distance at 250 ft, we find l, 6, 8, and 0 correlations in the four sections
reading from left to right, top to bottom. But the expected number of
correlations to be found in these four sections, if the correlations occurred
at random, are 4.2, 2.8, 4.8, 3.2. The value of chi-square for the observed and
expected values is equal to 11.4. There is only 1 degree of freedom in this 2 x
2 table, and the probability that this large value of chi-square could occur at
random is less than 1%. Therefore, there is little doubt that the distance coordinate
is meaningful, and we conclude that the motorcycle was moving through Dealey
Plaza and did, in fact, detect the sounds of gunfire.
Looking at the
information in Table II and in Fig. 22, we can determine that at least 6 of the
15 correlations above the detection thresholds are false alarms. These six
false alarms are indicated in Fig. 22 with an X drawn over them, and they are:
l. The fourth entry
in Table II that occurred at 137.70 sec is a false alarm, because it represents
a rifle shot fired from the knoll at Target 4 near the triple underpass at a
time when the limousine was near the position seen in frame 171. Thus, this
shot was fired in a direction opposite to that of the logical target.
2. The entry in
Table II that occurred at 140.32 sec is a false alarm, because it occurred only
1.05 sec later than earlier correlations also obtained from the TSBD. The rifle
cannot be fired that rapidly. Since there are three correlations plausibly
indicating the earlier shot, the one occurring 1.05 sec later must be a false
alarm.
3. The fourth entry
in Table II that occurred at 139.27 sec is a false alarm, because the
motorcycle would have had to travel 130 ft in 1.6 sec (55 mph) to gain that
position.
4,5,6. The second
and third entries at 145.15 sec and the third entry at 145.61 sec are false
alarms, because the motorcycle would have had to travel at 16 mph to gain the
indicated position of only 70 ft behind the limousine at the time of the last
shot. The motorcycle noise level (see Fig. 4) decreased by about 10 dB Just 3
sec before the time of the first correlations, indicating a slowing to
negotiate the 120 deg. turn onto Elm St. The motorcycle noise level did not
increase for the next 13 sec, so it could not have increased speed to 16 mph
and maintained it.
There remain nine
correlations that exceeded the detection threshold, and they occur at four
different times:
Group 1.137.70 sec
-- four correlations with test shots from the TSBD at Targets 1 and 3.*
Group 2.139.27 sec
--three correlations with test shots from the TSBD at Target 3.
Group 3.145.15 sec
-- one correlation with a test shot from the knoll at Target 3.
Group 4.145.61 sec
-- two correlations with test shots from the TSBD at Targets 3 and 4.
There is no other
acoustical evidence that would help to determine which of the remaining nine
correlations are false alarms, if any. Clearly, at least one of the first two
groups of correlations and at least one of the second two groups of
correlations must contain detections, because the order found in the data would
not likely have occurred by chance. The probability that two detections have
been achieved and that one is near 138 sec and the other near 145 sec is at
least 95%.
However, the
expected number of false alarms to be found when testing four different impulse
patterns is 13 (see Appendix C), and only six have been found. Therefore, it is
not unreasonable to expect that there are seven more, although that would be
the largest number possible since at least two of the remaining nine are probably
detections. The best that can be safely assumed is that each of the nine
remaining correlations is equally likely to represent a detection or a false
alarm.
*Possibly
because of the presence of an overhead sign that interfered with test shots at
Target 2, no correlations were found with that target.
On the basis of
this judgment and the assumption that each of the 15 events are independent,
the probabilities of several different Outcomes can be calculated.
The probability
that at least two shots have been detected is 96%, the probability that at
least three shots have been detected is 75%, and the probability that four
shots have been detected is 29%. The individual probabilities that shots
occurred at each of the four times at which correlations exceeded threshold are
88%, 88%, 50%, and 75%, listed in order of increasing time. The combined
probability that there were three shots and that the third (knoll) shot was one
of them is 47%.
Our correlation
detector that located the origin of gunfire also located the position of the
radio that transmitted the gunfire sounds. It is important to show that the
motorcycle trajectory determined by the detections is compatible with
independent evidence about a motorcycle trajectory. The necessary independent evidence
to show this compatibility is partially obtained from the positions of the
Presidential limousine and a motorcycle shown in the movie taken by Hughes (see
footnote on p. 62). This movie shows the limousine just turning Onto Elm St.
just before a motorcycle passes that has turned onto Houston St. from Main St.
We estimate that the motorcycle was at point M (Fig. 12) at that sighting. We
estimate that the limousine was at the position of microphone 2(9) (Fig. 12) at
that sighting, 215 ft north on Houston St.
The position of the
limousine at the instant of the President's head wound is shown in Fig. 22 at
two different times, assuming that either the third or the fourth shot struck.
Photogrammetric determination of the limousine speed on Elm St. was about 11
mph. The limousine's positions at times before the head wound is shown by the
two parallel lines projected backward, having a slope equal to 11 mph. The two
times at which the limousine position is equal to its assumed position when the
motorcycle was at point M are shown in Fig. 22. We find that these times were either 6.5 sec or 7.2 sec before
the first shot was fired. The motorcycle position at either one of these two
times was 180 ft away from its position when the first shot was fired,
according to the results of our correlation detector. Therefore, its average
speed north On Houston St. would be either 15.9 mph or 18.6 mph, depending upon
whether the third or fourth shot caused the head wound. These two trajectories
are shown in Fig. 22 also.
A precise
motorcycle location at the time of the third shot, calculated by Weiss and
Aschkenasy, was found to be 5 ft southwest of microphone position 3(4). This
location is marked in Fig. 22. The straight line that passes through this
point, and best fits the eight other microphone locations that produced echo
patterns indicating the other three shots, plotted in Fig. 22. This line is the
estimated motorcycle trajectory on Elm St., and it indicates an average speed
of 10.6 mph.
The complete
motorcycle trajectory shows that the motorcycle traveled north on Houston St.
at about 17 mph. It slowed to about 10 mph at a point about 40 ft south of the
corner at Elm St., and then continued west on Elm St. at about 10 mph. This
single diminution of speed is compatible with the single diminution of
motorcycle noise about 3 sec before the first shot is heard (see Fig. 6). We
conclude that the motorcycle trajectory determined by the gunfire detections is
compatible both with the positions of a motorcycle shown in the Hughes film and
With the loudness of the motorcycle noise as transmitted to the Dispatcher.
6. ADDITIONAL
RELEVANT SOUNDS ON THE DPD CHANNEL 1 TAPE
In an attempt to
gain as much acoustical evidence as possible, the Channel l tape was examined
for other relevant sounds. These other sounds consisted primarily of the
tolling of a bell, the noise of sirens, and voice and other transmissions.
6.1 Bell
The toll of a bell
can be heard faintly at about 152.5 sec. It was hoped that the location of the
bell, and therefore of the radio transmitter, could be obtained by acoustically
identifying the bell.
The energy spectrum
of the 1/3-sec segment containing the bell sound is shown in Fig. 23. Several
peaks evident in the spectrum are harmonically related. The fundamental
frequency of this series of spectral peaks is 210 Hz. The spectral peaks are
marked according to the usual nomenclature used to describe overtones of a
carillon bell. The fundamental tone is called the hum note. The second
harmonic, called the strike note, is at the nominal pitch of the bell -- in
this case, 420 Hz. The third harmonic is a fifth above the strike note. Higher
harmonics are strong at 1050 Hz and 1470 Hz. The minor third above the strike
note is strong, and this fact is characteristic of carillon bells.
The tape-recording
system was found to be about 5% slow, when the time annotations were measured
with a stopwatch (see Fig. 9). Therefore, the apparent pitch of the tone would
have a frequency of (1.05) (420) = 441 Hz.
FIG. 23. ENERGY
SPECTRUM OF TAPE SEGMENT CONTAINING THE SOUND OF A BELL.
Careful
investigation by the Committee staff did not discover any such bell within
earshot of Dealey Plaza. During the acoustical reconstruction tests in Dealey
Plaza, the sounds of railroad locomotive bells were recorded and subsequently
analyzed. These sounds bore no similarity to the carillon-like sounds of the
original recording.
We concluded that
the bell sound on the Channel 1 tape recording must contain sounds from at
least one transmitter not in Dealey Plaza at a time near 152.5 sec.
6.2 Sirens
The region from 263
to 300 sec of the stuck transmission contains the sounds of a number of sirens.
The effect is not that of a microphone being carried on a vehicle with a
wailing siren, but rather of many vehicles with sirens coming and going around
the microphone.
6.3 Voice and Other
Remote Transmissions
Starting just after
264 sec, a voice transmission says, "Anybody know where 56 is?" The
quality of this voice is such that it sounds as if it may have been picked up
by the open microphone of the stuck transmitter, rather than having come from a
second transmitter on the same channel, but it is impossible to tell for sure.
In many other
cases, there are brief voice signals from other remote transmitters. Sometimes
these signals are too faint to be understood (such as the voice signal shown in
the spectrograms in Fig. 5), sometimes they are loud but very distorted, and
sometimes they are quite intelligible. These competing transmissions are often,
but not always, accompanied by heterodynes, which are tones caused by slight
differences in frequency among the competing transmitters. Many times these
remote transmissions are very brief (around O.l sec) "beeps" with no
voice, signifying attempts to make one's desire to use the channel known. This
beeping is common practice on a shared radio channel.
7. REVIEW OF AN
INDEPENDENT ANALYSIS OF THE POSSIBLE THIRD SHOT
Owing to the
uncertainty about the possible third shot found in our study, the Committee
sought an independent analysis. Professor Mark Weiss and Mr. Ernest Aschkenasy
of Queens College conceived of an analytical extension to our work that could
determine with more certainty whether or not the match between one echo pattern
from our acoustical reconstruction with one impulse pattern on the DPD tape
indicated a third shot. At a meeting on October 24, we contributed to the
design of this analytical work.
Their analysis was
conducted as follows. First, they made a graph of the waveform of the echo
pattern we recorded on microphone 3(4), when a rifle was fired from the knoll
at target no. 3 (see Table II). From this graph, they identified the 22 loudest
individual echoes within the pattern. Then, they identified the 22
echo-producing objects within Dealey Plaza by noting which objects corresponded
to observed echo delay times - i.e., by identifying
rifle-to-object-to-microphone sound paths that would account for the times each
of the 22 echoes were received by microphone 3(4).
Next, they
analytically moved the position of microphone 3(4) several times by calculating
for each time what the echo pattern would have looked like if that microphone
used in the acoustical reconstruction had been located in these other
positions. After a time, they found that a position about 5 ft southwest of the
actual location of microphone 3(4) represented the true location of the
motorcycle at the instant the muzzle blast would have been received by its
radio. Then they calculated the delay times for each of the 22 echoes received
at that point as it moved down Elm St. at 11 mph. The resulting series of
echoes was found to match with the sound impulses on the DPD tape beginning at
about 144.9 sec (see Table II).
Weiss and
Aschkenasy found that 12 Of the 22 echoes were loud enough to exceed a
threshold that they felt excluded most of the weak echoes that would not be
audible in the DPD tape. They found that 10 of these 12 echoes occurred within
-+1 msec of the occurrences of 10 of the 14 impulses on the DPD tape that were
loud enough to exceed a threshold. The value of the correlation coefficient
that represents this match is 0.77. This value exceeds the threshold value of
0.60 for which we accept a correlation as possibly indicating a shot.
The probability
that a false alarm will be produced by the correlation scheme used by Weiss and
Aschkenasy is much lower than it is by our correlation scheme, because in our
analysis we counted echoes that occurred within +/- 6 msec of the occurrence of
impulses on the DPD tape. We were required to count echoes occurring within
this larger time interval, because of our initial uncertainty of the true
motorcycle location.
We computed the
probability that Weiss and Aschkenasy could have obtained by chance their good
match between their calculated echo pattern and the impulse pattern on the DPD
tape. We observed that they obtained 10 matches, to a precision of +/- 1 msec,
out of 12 test echoes, with 14 impulses in a 320 msec time span. We note,
however, that the 12 test echoes were contained in two time intervals of
90-msec total duration. These two intervals were separated by a span of about
230 msec in which no echoes appeared. Because an echo was counted if it
occurred within a 2-msec time window, there were 45 possible windows in which
echoes may occur. Since one of the 10 occurrences can always be matched simply
by adjusting the origin of the time scale, there are only 9 independent
occurrences. The probability of obtaining by chance 9 or more out of 12 echoes
occurring within any specific 14 time windows out of a possible 45 is equal to
3.13x10(')-4. This probability of obtaining by chance as good a match as was
obtained on a single try is given by the hypergeometric probability function.
However, they were required to try not once, but about 180 times. This is
because the motorcycle could have been anywhere in a 40-ft by 18-ft rectangular
space. Since a significantly different pattern would be calculated by them for
each different 2-ft by 2-ft square, they were required to examine about 180
different patterns. The probability of obtaining )Just one match by chance in
any of 180 independent tries is equal to 5.3x10(')-2, or about 5%. Therefore, the
probability that they obtained their match because the two matched patterns
were due to the same source (gunfire from the knoll) is about 95%.
APPENDIX A.
COMPUTER SIGNAL PROCESSING
Many of the
analyses of the acoustic data were performed on digital computers. In this
appendix, we describe these processing methods.
A.1 Digitizing
When played from a
magnetic tape, sound is in the form of a continuous electrical signal. For it
to be amenable to processing by a digital computer, its voltage must be sampled,
or read, at frequent intervals. The voltage must then be expressed as a digital
quantity. The sampling rate must be sufficiently rapid to preserve the
high-frequency components of the signal; sampling rates of 10,000 times per sec
and 20,000 times per sec were used in this work. The signal must then be
digitized with an analog-to-digital converter; the resulting series of numbers
is stored on a computer disk file.
A.2 Interactive
Playback and Display
Once the signals
have been digitized, waveforms can be graphically plotted on a computer
display; the signals may also be reconverted to sound by a digital-to-analog
converter. Interactive signal display, editing, and playback programs make it
possible to display any time interval of the signal and to convert it back to a
sound signal for listening. This interactive process of observing portions of
the signal waveform and simultaneously listening to it is very valuable.
A.3 Plotting
In addition to
showing portions of the signal waveform on the computer display, we also used
the computer and a graphical plotter to make pen-and-paper drawings of the
signal waveforms. These high resolution plots, usually made with a scale of 5
in. per 1/10 sec, provide a permanent record of the signal. Examples of these
plots are shown in Sec. 3 of this report.
A.4 Signal
Enhancement
Computations
performed on the digitized signal can produce filtered versions and other
representations of the signals. Digital signal processing can accomplish the
same kinds of filtering that can be performed in the analog domain, and it can
accomplish new kinds of filtering that are impossible by conventional means.
Several different kinds of signal processing were performed on the data.
Enhancement by
Adaptive Noise-Canceling Filter
An adaptive
noise-canceling filter differs from fixed filters in that it automatically
adjusts its signal-processing characteristics by means of an algorithm that
allows it to predict certain noise components. The particular filtering
process* used for the Channel 1 tape allows the filter to separate periodic
components of the noise from random components. Periodic components are those
elements of an input signal that repeat at regular intervals -- for example,
the ticks of a clock and a 60-Hz powerline voltage hum.
<*Widrow,
B., Glover, J.R., Jr., McCool, J.M., Kaunitz, J., Williams, C.S., Hearn, R.H.,
Zeidler, J.R., Dong, E., Jr., and Goodlin, R.C., "Adaptive Noise
Canceling: Principles and Applications," Proc. IEEE 63, 1692-1716
(December 1975).
One property of
periodic components is that, given sufficient past history, they can be
predicted; indeed, a perfectly periodic signal can be predicted perfectly. The
filter "learns" from the past history of the signal, estimates the
signal for the next time period, and subtracts its estimate from the input.
What is left are those portions of the signal that the filter cannot estimate
-- i.e., the random components.
A time delay was
inserted into the processing system, Just ahead of the adaptive filter, to
assist in controlling the separation of periodic and random components. Random
components having time duration less than that of the time delay pass through
essentially unaffected by the filter. These random components form the primary
output of the filter. A second output was the periodic component that was being
subtracted out; this subtracted information was also saved in digital form on
disk. Examination of this subtracted signal, by aural and visual means, yields
considerable insight into filter performance. Several test signals were fed
into the filter to verify proper operation and to adjust the various filter
parameters. The filter performed very well on the various test signals.
On the DPD Channel
1 tape, anticipated periodic and undesirable interferences included components
of motorcycle cylinder firing, powerline hum, heterodyne "squeals.,"
and occasional speech. Sections of this tape were played into the filter with a
wide range of filter parameter values. Filter action was monitored by listening
to both the primary and the secondary outputs. The filter removed residual
powerline hum, some speech, and heterodyne "squeals" of time duration
longer than that of the time delay. However, it accomplished little with
respect to what had been believed to be motorcycle noise. We therefore
performed an autocorrelation analysis, as described below.
Autocorrelation
Analysis of Motorcycle Engine Noise
Our interpretation
of the sounds on the Channel 1 tape would have been made much easier if we had
had some knowledge of the movements of the motorcycle carrying the microphone.
For example, if we had had information on when the motorcycle was moving
steadily (along a straight street), slowing down and possibly shifting gears to
turn a corner, or stopping, we might have been able to infer whether these
movements were consistent with travel into or through Dealy Plaza. However, we
did not have this information. Thus, to determine the engine speed with greater
accuracy than is possible from engine loudness, we wrote a computer program
that would compute the short time autocorrelation function of the motorcycle
noise signal. This function assesses the similarity of a signal with itself
shifted in time; if the signal is periodic, this similarity will peak when the
signal is shifted by one period.
This
autocorrelation analysis program was applied to the stuck transmission period
on the Channel 1 tape. The results showed no periodicity that we could
attribute to motorcycle engine firing. As a test case, this program was also
applied to a high-fidelity recording of motorcycle engine noise, and it clearly
showed the known periodicity of the test signal. Although our failure to detect
the motorcycle engine periodicity is puzzling, it is consistent with our
inability to perceive the engine firing clearly when we are listening to the
tape, and it is also somewhat consistent with the failure of the adaptive
noise-canceling filter to filter out a coherent motorcycle engine sound signal.
Enhancement by
Spectral Subtraction
A third method applied
to enhance the Channel 1 signals was the subtraction of a noise spectrum
estimate. This method is currently under development at BBN, under U.S.
Government sponsorship, for the enhancement of speech signals in the presence
of stationary flat-spectrum additive noise.* It is similar to, but somewhat
more general than, the INTEL enhancement method developed by Weiss et al. We could not tell whether this method would
be effective with non-stationary non-flatspectrum noise, but since the program
was already available, we tried it.
In this method, the
signal is converted by a Discrete Fourier Transform to a magnitude spectrum and
a phase spectrum. A previously computed estimated noise spectrum is subtracted
from the magnitude spectrum; the altered magnitude spectrum is then recombined
with the phase spectrum converted back to a waveform by an Inverse Discrete
Fourier Transform. Several parameter settings for this filtering method were
used with a portion of the Channel 1 tape. None were successful in reducing the
motorcycle noise without introducing noise transients attributable to the
filtering process.
*Berouti, M.,
Schwartz, R., and Makhoul, J., "Enhancement of Speech Corrupted by
Acoustic Noise," IEEE Int. Conf. on Acoustics, Speech, and Signal Processing,
Washington, DC, April 2-4, 1979.
+Weiss, M.R., Aschkenasy, E., and Parsons, T.W., "Study and Development of
the INTEL Technique for Improving Speech Intelligibility," Nicelet
Scientific Corp., Report RADC-TR75-108, 1975.
APPENDIX B. RADIO
TRANSMISSION OF GUNFIRE SIGNALS
The 1963 DPD
Channel 1 radio link and recording system contained the following components:
microphones, radio transmitters, an RCA Fleetline radio receiver (Model
C9F350), and a Dictabelt recorder. Radio systems such as this are designed to
carry speech signals and therefore incorporate signal modifiers to optimize the
dynamic range and bandwidth of the system with respect to voice transmissions.
Since these signal modifiers are usually incorporated in the transmitter,
rather than in the receiver or the recording device, we focused our efforts to
simulate the radio link on the transmitter/microphone combination.
Among the radio
transmitters in use by the DPD in 1963, House Committee researchers found that
five different models were used on motorcycles. These were
* Motorola Model
FMT-41
*Motorola Model T-31BAT
*Motorola Model U-41GGT
*Motorola Model T-41GGT
*General Electric Model MT-13-N.
At the time of this
study, it was very difficult to find manuals for these models and even more
difficult to obtain access to a working unit. With the manuals-we were able to
find and with assistance from Motorola factory personnel, we discovered that
the microphone used with the T-31BAT would have been Motorola NMN 6006A and
that microphones used with the Other Motorola transmitters would have had
similar characteristics; i.e., they would have been dynamic cardioid types with
internal preamplifiers. We have no information about the GE radio model and its
microphone. We eventually located a T-3lBAT owned by the Boston Metropolitan
District Commission Police Department. The MDC kindly made this radio and a GE
Model ERSlA receiver available to us.
The basic scheme
used in this and other Motorola radio transmitter/microphone systems of the
same vintage is sketched below. This type of circuit limits the slope of the
audio signal rather than its amplitude. Therefore, it will limit high frequency
signals more than low-frequency signals, as shown in Fig. 11 of this report.
The frequency response of the system rolls off at 36 dB/octave above 3 kHz and
at 6 dB/octave below 2.3 kHz. The signal, in effect, is differentiated and
low-pass filtered. The smoothed, calculated frequency response of the system is
plotted in Fig. B.1.
Our procedure for
obtaining the data shown in Figs. 10 and 11 was to play tape recordings of
gunfire, made anechoic by time gating, through a circuit designed to simulate
the frequency response and amplitude-limiting characteristics of the Motorola
6006A microphone into a second tape recorder. We then took the second tape to
the MDC Police radio shop. There, we played this tape through a variable
attenuator (to control the level of the signal being put into the transmitter),
through the Motorola transmitter, through the GE receiver, and onto another
tape recorder. This third tape recording was played back into an oscilloscope
and photographed producing the waveforms shown in Fig. 10. Peak-to-peak
amplitudes of these waveforms were measured and plotted to produce Fig. 11.
FIG. B.1
TRANSMITTER/MICROPHONE SYSTEMS FREQUENCY RESPONSE.
In addition to
having had similar effects on the waveforms recorded on Channel 1, the DPD
recording shows evidence of a time constant in the O.1 to 1.0 sec range. This
AGC does not occur in any of the Motorola transmitters. It could, therefore,
have been caused by the GE transmitter, by the receiver, or by the recorder.
APPENDIX C.
ANALYSIS OF FALSE ALARMS IN THE CORRELATION DETECTION TEST
The process of
binary correlation that was used to detect gunfire echo patterns among the
impulse patterns on the DPD tape can, like any other detector, produce false
alarms. This analysis determines the number of false alarms to be expected from
random noise impulses on the DPD tape.
Each echo pattern
contains an average of M = 12 echoes in a 1/2-sec span. But, we consider each
echo to have a +/- 6-msec acceptance window to account for echo time
differences introduced by not knowing the motorcycle position relative to the
test microphone positions. Therefore, there are about N = 40 different time
slots in which the 12 echoes may exist.
Each impulse
pattern contains some number of impulses ranging from n = 8 to n = 17, also in a
1/2-sec span.
The matching
process seeks to find the number of impulses, that lie within the acceptance
windows of the echoes that comprise the echo pattern. If the impulses are
caused by a random noise source, then the number of matches, i, is what would
be expected from random sampling n times a population of N that contains M
echoes. The probability of getting I matches at random is given by the
hypergeometric probability distribution p(N, M, n, i).
The correlation
coefficient is defined to be equal to i/mn. The probability of obtaining a
correlation coefficient equal to 0.6 or greater was calculated for N = 40, M =
12, and n = 8, 10, 12, 13, 14, 17. The results for the six successive values of
n were: 4.8x10-3, 6.0x10(-3), 8.5x10(-3), 1.0x10(-2), 1.2x10(-2), 1.5x10(-2).
For an impulse
pattern having 10 impulses (n=10), there are expected (6.0x10-3x432) = 2.6
false alarms, because there ave 432 echo patterns to correlate with. There were
four impulse patterns that were correlated with all 432 echo patterns, and they
had n = 8, 10, 12, 17 impulses on them. The total number of false alarms to be
expected works out to 13.
This number was
Judged to be acceptably small, so the detection threshold value was set at 0.6.