Impulsive Behavior:
The CourtTV -
Sensimetrics Acoustical Evidence Study.
D.B. Thomas
(C) 2003 Donald B. Thomas
(Printed here with the permission of the author. All rights reserved)
"Impulsive - characterized by undue haste and lack of thought or deliberation."
- Webster's Dictionary
In November 2003 the CourtTV
program Forensic Files reported that a new study of the acoustical
evidence commissioned by them with a signal analysis firm, Sensimetrics Inc.,
had concluded that there is no valid evidence for gunshots on the Dallas Police
Department (hereinafter DPD) recordings, contrary to the findings of the House
Select Committee on Assassinations (hereinafter HSCA) in 1978. According to the
Sensimetrics report (available at the CourtTV website),
the match between the suspect sound pattern on the DPD recording and a test
shot fired from the grassy knoll, was no greater than expected to occur by
chance.1
Sensimetrics wrote a computer program called Impulses which
automatically compares the waveform of the suspect sound identified as a grassy
knoll gunshot echo pattern, to the virtual pattern simulated by HSCA acoustical
experts in 1978 (the Impulses program is also available for downloading
at the CourtTV website). The actual test shot pattern from the
grassy knoll is not directly comparable to the suspect pattern because it was
recorded in August 1978 when the air temperature was 90
degrees. The assassination occurred in November when the air temperature was 65
degrees. The comparison of the patterns is based on the arrival time of the echoes
at the microphone and the speed of sound is affected by air temperature. Also,
the test shot was recorded on a stationary microphone, whereas the working
hypothesis holds that the suspect sound was recorded on a police motorcycle
that was in motion traveling with the president's motorcade. The HSCA
consultants adjusted for these factors to simulate a virtual echo pattern that
could be compared to the suspect sound on the police recording. Moreover, they
had to simulate 180 virtual patterns to cover a (2' x 2') grid of potential
motorcycle positions for an 18 x 40 ft area surrounding the test microphone
position that recorded the actual grassy knoll test shot. They found their best
match with a virtual pattern for a position about 5 ft SW of the test microphone
position.2 The HSCA consultants found
that when the muzzle blast impulse from the virtual pattern is aligned with the
leading impulse on the DPD pattern, the result is 10 coincidences among the 14
succeeding impulses and the 12 known echoes in the test shot pattern
(corresponding impulses were scored as coincident if they were within 1 msec of
one another).
The HSCA team believed that
the two extraneous impulses on the DPD recording could be explained as stray
static (the static marking rate was ca. 8 per sec, and the duration of the
pattern is 370 msec). The two echoes that did not score a match with the DPD
pattern originated with structures near the grassy knoll (the building where
Zapruder was standing) and did match to two smaller impulses on the DPD
pattern. It was theorized that because the motorcycle microphone was behind the
windshield, echoes coming from the front might have been attenuated. Even so,
by scoring those two as "missing," the match of 14:12:10 is highly
statistically significant (unlikely to arise by chance).3
However, Sensimetrics
reported that in making the same comparison, applying "less
subjective" criteria, only five coincidences are found. By comparing the
simulated grassy knoll pattern to other parts of the DPD recording, around five
matches also appear according to the Sensimetrics report. They therefore
concluded that the DPD suspect pattern resembles a grassy knoll gunshot no more
than would be expected by chance.
Figure 1
Figure 2
It turns out that the Sensimetrics
analysis contained serious errors. I was able to find the match reported by the
HSCA with the Impulses program in short order. Sensimetrics had
applied the wrong criteria.
As Indiana Jones said in Raiders
of the Lost Ark:
"They're
looking in the wrong place!"
The first problem with the Sensimetrics
analysis was they used a different playback of the DPD dictabelt than the one
used by the HSCA experts. There are at least three different playback versions
of the Dictabelt in existence. Jim Bowles of the DPD made one back in 1963
using a rented dictaphone machine. The second was made by the HSCA consultants
in 1978 in order to authenticate the Bowles recording. The HSCA consultants
used the Bowles recording for their analyses, instead of their own playback,
because it was considered to be the more reliable record of the sounds as
recorded in November 1963, due to the wear and tear experienced by the original
dictabelt over the intervening years. The third playback was made by the NRC
panel in 1981 using the FBI's equipment. The waveform that appears in the Impulses
program, identified as the HSCA.wav file, seems to be one of the latter. One
can see differences in the relative amplitudes of some of the impulses between
the Bowles pattern analyzed by the HSCA and the Sensimetrics version.
Both patterns are attached here. In theory it should not make any difference
which version of the pattern is analyzed, but in practice it does because the
different playbacks introduced different anomalies as discussed below.
The first important anomaly
involves the playback speed problem. In making his copy, Bowles rented a
dictaphone machine to playback the police dictabelt. Because they are designed
for dictation, there is a dial that can be used to adjust playback speed. But,
because there is no standard speed setting, the use of different machines made
it inevitable that a speed warp would creep in. By analyzing the 60 Hz power
hum the HSCA consultants found that the rental machine was playing about 5%
faster than the original dictaphone machine when it recorded. Therefore, time
is compressed on the Bowles tape by about 5%.4 For example, the NRC panel found that the time
interval between the broadcast by Sheriff Decker and a later broadcast by Sgt.
Bellah was 171 sec on the Bowles tape, but the same interval is 178 sec on the
FBI version.5 The time intervals can
be decompressed by using a speed correction factor of 1.05 if one uses the
Bowles version. The HSCA consultants achieved their match by tweaking the
correction factor to 1.043 (as did the NRC panel when they used a spectrograph
to match the adjacent Decker speech patterns, more on this later).
The mistake made in the Sensimetrics
analysis was to use the wrong speed correction factor. They used 0.95, which
compressed the signal further instead of decompressing. When I first tried the Impulses
program I applied the 1.05 and 1.043 correction factors and got even fewer
coincidences (3). But noting the difference in the amplitude of some of the
impulses I went back and ran the program with no correction (that is, a speed
setting of 1.0). That produced the match found by the HSCA. The HSCA and FBI
recordings were speed adjusted using the 60 Hz power hum and therefore do not
require a speed correction. My colleague Michael O'Dell compared the sound clip
packaged with the Impulses program and confirms that it is identical in
speed to the NRC's dictabelt playback.
The third mistake made by Sensimetrics
was to use the wrong signal to noise ratio in establishing the threshold for
inclusion of impulses for comparison. In their first comparison of the test
shot and the DPD suspect pattern, the HSCA team did not apply a threshold. They
wanted to find the best alignment between their 180 simulated test patterns,
each with 26 echoes, and the impulses on the suspect pattern. A perfect
alignment was achieved with a simulation based on a microphone position about 5
ft away from the test microphone position. Such a high degree of match is
deceptive because there are so many low amplitude impulses and this greatly
increases the odds of getting a spurious match (a false positive). Because most
of the low amplitude impulses on the DPD recording are motorcycle piston
firings, a threshold has to be applied to separate the suspect signals from the
motor noise. The threshold applied by the HSCA analysts was based on the noise
level in the segment of recording immediately prior to the impulses assumed to
be the shock wave and muzzle blast. This S/N ratio left 14 impulses above the
threshold and 12 impulses of similar amplitude on the test shot pattern. Sensimetrics
misunderstood the logic of this approach and used a threshold that left 26
impulses on the DPD pattern for comparison.
By applying the correct
speed factor (1.0) and the correct threshold with 14 impulses (a setting of
6450 units) the Impulses program found 8 coincidences (Fig. 1). By
clicking on the red dots on the Impulses display, a readout of the
impulse position (in msec) appears, and the echo delay time can be extrapolated
by subtracting the position of the putative muzzle blast, which is also displayed.
These values can then be compared to the echo delay times and pathways in Table
4 of the HSCA report. Those pathways and corresponding echo delay times for the
matching impulses are: (4)19,
(14)71, (15)78, (18)279, (19)284, (20)285, (22)311, & (23)312
msec. Obvious question; what
happened to the other two coincidences reported by the HSCA? These can be
accounted for by the anomalies that still exist in the Sensimetrics
program.
Note in the Bowles version
of the DPD pattern (Fig. 2) there is an impulse numbered 16 which is much
larger than the adjacent impulses numbered 14 and 15. It is one of the largest
two or three impulses in the whole pattern. In the Sensimetrics
waveform, the same impulse falls below the S/N threshold, now much lower than
the adjacent impulses. By lowering the S/N threshold to capture 18 impulses (a
setting of around 5800), this impulse now matches to one of the echo paths
defined by the HSCA study (No.16 in their table 4). This was probably the ninth
coincidence identified by the HSCA.
This brings us to yet
another error by Sensimetrics. They only entered 25 of the 26 echo delay
times determined by the HSCA consultants. The one left out was the last, an
impulse that arrives at 369 msec after the muzzle blast (on the Impulses
program the virtual echo positions are indicated by black dots on the S/N
threshold setting). This is the large amplitude impulse identified as No. 26 in
the DPD suspect pattern. Because the Impulses program does not include
this echo from the test shot it couldn't match up and there is no way to say
for sure whether it would or would not match up to within 1 msec with this
impulse No. 26 had it been included. I suspect that this was the tenth
coincidence found in the HSCA study.
Even discounting for these
anomalies, the 8 coincidences that are found by the Impulses program is
statistically significant ( the program calculates that p = 5.5 x 10-3).
Hence, there is an objective match between the suspect DPD pattern and the
grassy knoll test shot even though Sensimetrics failed to find it. The Sensimetrics
and CourtTV folks also failed to understand that the matching
data in and of itself does not establish evidence for gunfire on the
recordings. It was the order in the matching data that compelled the acoustical
experts, and the HSCA, to conclude that the gunfire is on the recordings.
CourtTV did not address that evidence.
A concept that they failed
to grasp is basic. In forensic testing one is comparing an observed result to
an expected. If the observed matches the expected, the result is positive. If
they don't match, the result is negative. In this instance it should have been
realized that any small error in the procedure would result in a false
negative. But no error, large or small, objective or subjective, can produce a
false positive. Only chance can produce a false positive. If they had
understood that concept they would not have stopped looking when they failed to
duplicate the HSCA result. They needed to explain how the HSCA analysts got a
positive result, and if they thought the HSCA had gotten a false positive, then
they needed to explain, based on the real data and actual methods, why this
might be so. The Sensimetrics approach was indicative of a
disinclination for serious analysis. Their skills as signal analysis processors
and computer programmers are confirmed by the fact that one can achieve, in
essence, the same match found by the real acoustical experts in 1978. A lack of
familiarity with the evidence in the case, and perhaps a lack of time, resulted
in a lackadaisical study.
A further example of
analytical failure was the Sensimetrics approach to the cross-talk
problem. In his published transcripts of the DPD recordings, Communications
Supervisor James C. Bowles explained how broadcasts originating on one channel
could cross over to the other. He cited several instances of cross talk, one of
which was a broadcast by Deputy Chief Fisher. Bowles states:
“At approximately 12:31 (Channel 2), Deputy Chief
Fisher #4 discussed a traffic problem with Captain #125. The last sentence in
that message, “I’ll check it,”
Sensimetrics disputes Bowles identification saying on page 29 of
their report,
“It is not difficult to show by spectrographic
analysis that the two recordings cited contain two entirely different messages.
On the Channel 1 recording, the words are, “I’ll get it.” On Channel 2 a longer message
concludes, “I’ll check it.” The acoustic content of the two transmissions can
be seen to be very different (see Figure 19).
The question one is
compelled to ask is, if it is not difficult to show, then why not show us? The
presentation of an unanalyzed spectrogram is not a scientific argument. It’s as
if a fingerprint were introduced into evidence without the points of identity
used to objectively determine the match, or lack thereof. Instead the reader is
invited to examine the patterns visually and reach the suggested conclusion.
There is no explanation of how Sensimetrics determined that the middle
word is “get,” and not “check” even though the ch- sound is
obvious to this listener as it was to Bowles. Some dissenters, including the
NRC panel, believe the vocalization is, “All right Chaney.” Does a
spectrograph “speak” for itself, or does it require interpretation? It is not
difficult to show that the proper interpretation of spectrographs requires more
than cursory examination. Consider the experience with another instance of
cross talk, the "Double Decker." In a separate passage on page 25 of
their report, Sensimetrics makes reference to the Decker cross talk
declaring,
“There is no ambiguity regarding the content of the
voice transmission.
But it was the “other”
results that demonstrated the identity of the broadcast. The NRC’s
spectrographs suggested the opposite. Because it was important to the NRC panel
to demonstrate that the barely audible speech transmission on Ch-1 was
identical with Sheriff Decker’s orders to surround the Grassy Knoll and “hold
everything secure,” chairman Norman Ramsey arranged for the FBI signal analysis
unit to prepare voice prints of the corresponding broadcasts. Based on his
eyeball inspection of the spectrographs Ramsey declared the broadcasts to be
identical.7 But another panel member,
Richard Garwin, in a memoir of the panels operations knew that,
“Although visual inspection showed some similarities,
such an
Therefore the IBM computer
was instructed to vary the speed of one of the tapes in order to seek a match.
The computer found a match with a speed correction factor of .957. However,
the match was not a very good match. The correlation coefficient between the
two broadcasts at the optimum juxtaposition achieved by the computer was only
0.5. The NRC panel opined that the noisy background, which is different on the
two recordings, could be to blame. Therefore, in order to have a control or
check for their test, the computer was asked to compare another instance of
cross talk that was more audible than the Decker speech on Ch-1. They used the
aforementioned Bellah cross talk, “You want me … Stemmons.” Here the computer found a much more
robust correlation coefficient of 0.8.
Did the objective,
computerized comparison prove that the vocalization on Ch-1 was from Decker’s
broadcast on Ch-2? Perfect identity in correlation is a score of 1.0. A
correlation coefficient of 0.5 demonstrates a scientifically significant match
between the broadcasts, but not identity. If one accepts the reasonable
argument that the difference in the background noises contributed to the lower
score, then the 0.8 value achieved with the Bellah cross talk is arguably a
score demonstrating identity. A score of 0.5 does not establish identity, but
only a high degree of similarity. The reality is that these are not clear
channel broadcasts of the type for which a voice print analysis would be
appropriate. These barely audible fragments of speech are embedded in a noisy
background. There are sophisticated computer programs designed to "lift
and separate" to borrow an expression, but these have not been applied to
the DPD cross talks.
The reasons for accepting
that the Decker broadcasts are identical are two: 1) because the vocalizations
are attended by brief, loud tones called heterodynes, caused by policemen keying
their microphones, and these are present in the same number and pattern
bracketing the broadcasts on both channels, and 2) and most importantly,
because by listening carefully, one can hear the words “hold everything
secure.”
The vocalization “I’ll
check it,” is much clearer on Ch-1 than is the Decker broadcast. That is
why when lecturing on this subject I play the recording so the audience can
hear it for themselves. The Sensimetrics conclusion on this issue is
unattended by analysis and is based on nothing more than a biased opinion, and
perhaps a degree of hearing impairment. I encourage folks to listen to the
recordings and decide for themselves.
To listen to BOTH recordings (each repeated a second time)
Click HERE
(CAUTION: This is a very LARGE sound file
and not recommended for dial-up users)
1.- Berkowitz, R. 2003. Searching for Historic
Noise: a study of a sound recording made on the
2.- Weiss, M.R. & E. Aschkenasy. 1979. An
analysis of recorded sounds relating to the assassination of President John F.
Kennedy. Department of Computer Science, Queens College, New York.
3.- Thomas, D.B. 2001. Echo correlation
analysis and the acoustic evidence in the Kennedy assassination revisited.
Science & Justice 40: 21-32.
4.-
Barger, J.E., S.P. Robinson, E.C. Schmidt & J.J. Wolf. 1979.
Analysis of recorded sounds
5.- National Research Council, 1982. Report of
the Committee on Ballistic Acoustics. National
6.- Bowles, J.C. 1979. The Kennedy
Assassination Tapes: a rebuttal to the acoustical evidence
7.- Committee on Ballistic Acoustics, National
Research Council, 1982. Reexamination of
8.- Garwin, R.L. 1987. Examining the Kennedy
Assassination Evidence. Pp. 203-209, in, W.P.
Back to THE KENNEDY ASSASSINATION for the NOVICE