Still lower molecular weight compo-nents may be fractionated by high forcefield SFFF (12). For example, the single-stranded circular fD viral DNA macro-molecule was well retained at 30,000 rev/min (constant field), with peak retentioncorresponding to a molecular weight of1.71 x 106. This closely compares withthe reported value of about 1.7 x 106 forthis material (24). The protein fibrinogenwith a reported molecular weight of5 x 105 has also been retained with aretention ratio R = 0.37, representingthe approximate mass limit of separationwith our present apparatus (12).

Bacteria. By utlizing low SFFF forcefields it is feasible also to fractionaterelatively large biomasses, such as Esch-erichia coli bacteria (25).

References and Notes

1. J. C. Giddings, M. N. Myers, J. F. Moellmer, J.Chromatogr. 149, 501 (1978).

2. J. C. Giddings, S. R. Fisher, M. N. Myers, Am.Lab. 10, 15 (1978).

3. J. J. Kirkland, W. W. Yau, W. A. Doerner, J.W. Grant, Anal. Chem. 52, 1944 (1980).

4. M. N. Myers, K. A. Graff, J. C. Giddings, Nucl.Technol. 51, 147 (1980).

5. J. C. Giddings, M. N. Myers, K. D. Caldwell, S.R. Fisher, Methods Biochem. Anal. 26, 79(1980).

6. J. C. Giddings, F. J. F. Yang, M. N. Myers,Anal. Chem. 46, 1917 (1974).

7. G. Karaiskakis, M. N. Myers, K. D. Caldwell,J. C. Giddings, ibid. 53, 1314 (1981).

8. W. W. Yau and J. J. Kirkland, Sep. Sci. Tech-nol. 16, 577 (1981).

9. J. J. Kirkland and W. W. Yau, U.S. Patent4,285,810, 25 August 1981.

10. W. W. Yau and J. J. Kirkland, J. Chromatogr.218, 217 (1981).

11. J. J. Kirkland, S. W. Rementer, W. W. Yau,Anal. Chem. 53, 1730 (1981).

12. J. J. Kirkland, C. H. Dilks, Jr., W. W. Yau, inpreparation.

13. K. D. Caldwell, G. Karaiskakis, J. C. Giddings,J. Chromatogr. 215, 323 (1981).

14. C. D. Crowe and S. D. Abbott, personal com-munication.

15. J. J. Kirkland, W. W. Yau, F. C. Szoka, Science215, 2% (1982).

16. A. D. Bangham, D. de Gier, G. D. Freville,Chem. Phys. Lipids 1, 225 (1976).

17. K. D. Caldwell, G. Karaiskakis, J. C. Giddings,Colloid. Surf. 3, 233 (1981).

18. K. D. Caldwell, G. Karaiskakis, M. N. Myers,J. C. Giddings, J. Pharm. Sci. 70, 1315 (1981).

19. H. Gerber, M. Horisberger, H. Bauer, Infect.Immun. 7, 487 (1974).

20. H. Bauer, D. R. Farr, M. Horisberger, Arch.Microbiol. 97, 17 (1974).

21. M. Horisberger and J. Rosset, Experientia 32,998 (1976).

22. J. C. Giddings et al., Sep. Sci. 10, 133 (1975).23. K. D. Caldwell, T. T. Nguyen, J. C. Giddings,

H. M. Mazzone,J. Virol. Methods 1, 241 (1980).24. H. A. Sober, Ed., Handbook of Biochemistry,

(Chemical Rubber Co., ed. 2, Cleveland, Ohio,1970), p. H-4.

25. J. J. Kirkland and W. W. Yau, unpublishedstudies.

26. We thank Y-S. E. Cheng of Du Pont for theEscherichia coli bacteria, prepared as describedin G. H. Miller, Experiments in Molecular Ge-netics (Cold Spring Harbor Laboratory, ColdSpring Harbor, N.Y., 1972).

Reexamination of Acoustic Evidencein the Kennedy Assassination

Committee on Ballistic Acoustics, National Research Council

At the time of the assassination ofJohn F. Kennedy, a microphone, pre-sumably on a police motorcycle, wasstuck open and transmitted continuouslyon Dallas Police Department channel I.The transmissions were recorded on aDictaphone belt recorder, model A2TC.At the request of the House Select Com-mittee on Assassinations, this belt andmagnetic tape copies of it were studiedby J. Barger, S. Robinson, E. Schmidt,and J. Wolf(BRSW) of Bolt Beranek andNewman Inc., and later by M. Weiss andE. Aschkenasy (WA) of Queens College.In reports in September 1978 and Janu-ary 1979, BRSW concluded (1) that therecording contained four sounds, whichthey attributed to probable gunshots,and that with a probability of 50 percentone of the sounds (the third) was due to ashot from the grassy knoll area of DealeyPlaza in Dallas. Later, WA studied theecho patterns analytically, and their con-clusion (1) was that "the odds are lessthan 1 in 20 that the impulses and echoeswere not caused by a gunshot from thegrassy knoll, and at least 20 to I that theywere." BRSW subsequently reviewedthe results ofWA and concluded (1) that

SCIENCE, VOL. 218, 8 OCTOBER 1982

"the probability that they obtained theirmatch because the two matched patternswere due to the same source (gunfirefrom the knoll) is about 95%." Thisconclusion, together with the fact thatshots were definitely fired from anotherlocation, the Texas School Book Deposi-tory, was the basis of the finding by theHouse Select Committee on Assassina-tions that "scientific acoustical evidenceestablishes a high probability that twogunmen fired at President John F. Ken-nedy."

In response to a request- from theDepartment of Justice, the Committeeon Ballistic Acoustics was established bythe National Research Council in the fallof 1980 to review the methodology em-ployed in the evaluations of the recordedacoustic data and the conclusions about

the existence of a shot from the grassyknoll. The committee was also asked torecommend the kinds of tests, analyses,and evaluations needed to obtain betterinformation from the recording.

In the first months of its existence, thecommittee studied the analytical tech-niques used by BRSW-WA. Committeemembers found errors in the previousstudies which were sufficiently seriousthat, by the end of the first committeemeeting, no member was convinced byprevious acoustic analyses that therewas a grassy knoll shot.The committee continued its studies to

challenge its own conclusion and searchfor additional acoustic evidence. In par-ticular, it followed up a suggestion thatcross talk from Dallas Police Departmentchannel II was weakly recorded with thesounds attributed to gunfire on channelI. On the day of the assassination, chan-nel I was primarily used for normal po-lice activities. Channel II was used forthe presidential motorcade and was re-corded on a Gray Audograph disk. Thequality of the cross talk on the recordingfrom channel I was so poor that thecommittee could not conclude by listen-ing to the recordings from the two chan-nels that the two messages were thesame. Hence it made sound spectro-grams of portions of the two recordings.Analyses of the spectrograms showed

The Committee on Ballistic Acoustics includes Norman F. Ramsey, Harvard University, chairman; LuisW. Alvarez, Lawrence Berkeley Laboratory, University of California; Herman Chernoff, MassachusettsInstitute of Technology; Robert H. Dicke, Princeton University; Jerome I. Elkind, Xerox Palo Alto ResearchCenter; John C. Feggeler, Bell Laboratories, Holmdel, New Jersey, Richard L. Garwin, Thomas J. WatsonResearch Center, IBM Corporation, and Department of Physics, Columbia University; Paul Horowitz,Harvard University; Alfred Johnson, Bureau of Alcohol, Tobacco, and Firearms, National LaboratoryCenter, Department of the Treasury; Robert A. Phinney, Princeton University; Charles Rader, LincolnLaboratory, Massachusetts Institute of Technology; and F. Williams Sarles, Trisolar Corporation, Bedford,Massachusetts. Staff members at the National Research Council are C. K. Reed and Bertita E. Compton,Commission on Physical Sciences, Mathematics, and Resources. This article is based on the committee'sformal report, which will be released in November 1982.

0036-8075/82/1008-0127$01.00/0 Copyright © 1982 AAAS 127

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

conclusively that a segment of channel IIwas recorded on channel I at the same

location as the acoustic impulses attrib-uted to gunfire. From the channel II

recording it was clear that the message ofconcern was broadcast 1 minute after theassassination. This conclusion was con-

firmed by tracing time backward from a

later match of clearly intelligible cross

talk between the channels. In addition,features of the recorded sounds stronglysuggest that the open microphone was

not in Dealey Plaza at the time of theassassination.

of interest). Using the frequent annota-tions of time by dispatchers on bothchannels, BRSW estimated the approxi-mate time of the beginning and end ofthis 5.5-minute transmission. Because ofthe severe noise and distortion on therecording, the BRSW team could neitherconfirm that this segment contained gun-shot sounds, nor eliminate the possibilitythat they were present, by simply listen-ing or by examining the wave forms ofsounds on the tape. Therefore, they wentto Dealey Plaza on 27 August 1978 andmade recordings of test shots with vari-

Summary. Sounds recorded in Dallas at the time of the assassination of John F.Kennedy were analyzed by two research groups, whose reports formed the basis forthe opinion that two gunmen fired at President Kennedy. These reports and theacoustic evidence have been studied by the Committee on Ballistic Acoustics, andfurther acoustic analyses, including sound spectrograms, have been performed. Thecommittee finds that the acoustic data do not support a conclusion that a secondgunman was involved in the assassination.

For the reasons outlined above anddiscussed in more detail in this article,the Committee on Ballistic Acousticsunanimously concluded that:

1) Analyses of the acoustic evidencedo not demonstrate that there was a

grassy knoll shot, and in particular thereis no acoustic basis for the claim of 95percent probability of such a shot.

2) The acoustic impulses attributed togunshots were recorded about I minuteafter President Kennedy had been shotand the motorcade had been instructedto go to the hospital.

3) Therefore, reliable acoustic data donot support a conclusion that there was a

second gunman.

Description of Studies by

BRSW and WA

The acoustic studies by BRSW and thefurther acoustic studies by WA havebeen printed (1). This section is a simpli-fied presentation of the procedures as

described in their reports and as laterdiscussed when the committee met withBarger, Weiss, and Aschkenasy in 1981.The BRSW team began by listening to

tape copies of the recordings of bothpolice radio channels for general orienta-tion. Because the recorders were sound-activated, they could have stopped fre-quently, for varying amounts of time,except that the channel I recorder ran

continuously for 5.5 minutes when a

transmitter, presumably on a police mo-

torcycle, became stuck in the transmitmode (the channel I recorder appears tohave run continuously during the period

128

ous kinds of guns and ammunition, twoshooter locations, and many microphonelocations along the approximate route ofthe motorcade. For each combination ofshooter location and microphone loca-tion, there is a characteristic and com-

plex pattern of echoes in the recording ofthe test shot. Whereas the channel Irecording is noisy and distorted, the testshot recordings are clean records of theacoustical response of Dealey Plaza-except for some changes that took placebetween 1963 and 1978.The BRSW team then compared, man-

ually, each of 432 test shot wave formswith all the parts of the 5.5-minute rec-

ord that could reasonably have includedassassination gunshot sounds. They useda binary correlation metric, with a +6-millisecond window, applied to stripchart recordings of the relevant wave-

forms. For each suspected assassinationshot and each test shot, the strip chartrecordings were aligned for a best matchand a score was obtained by calculatingthe correlation coefficient. The correla-tion coefficient was defined as the num-

ber of impulses (large peaks) in the waveform of the suspected shot that came

within 6 milliseconds of an impulse inthe test shot, divided by the geometricmean of the numbers of impulses avail-able in the suspected shot and in the testshot. This correlation procedure doesnot make use of all the information avail-able (impulses that barely resemble eachother affect the score as much as im-pulses that match each other well), but itpermits relatively easy computation ofsimilarity. Only candidate shots thatgave a binary correlation greater than 0.6

were studied further. Fifteen pairs, in-volving only six sets of impulses onthe channel I recording, survived thisscreening.For each surviving pair, the location of

the microphone that recorded the testshot should approximate the location ofthe motorcycle at the time the suspectedassassination shot was fired. This time,in a relative sense, could be determinedby the location of the suspected impulsesalong the length of the tape. In a plot ofmicrophone position against suspectedshot time, a sequence of actual assassi-nation shots should lie along a line thatdescribes the movement of the motorcy-cle during the interval of the assassina-tion shots. Unrelated impulses (falsealarms) could lie on this trajectory orelsewhere, depending to some degree onchance. Most of the pairs could be iden-tified as false alarms in this way. Theremaining pairs were judged, by thecloseness of the trajectory fit, to containat least two assassination shots and atleast some false alarms. Based on aninterpretation of photographic evidence,the hypothesized motorcycle trajectorywas subsequently claimed to be consist-ent with that of the motorcycle driven byOfficer H. B. McLain in the Kennedymotorcade.Of the six sets of impulses that give

high binary correlation coefficients withtest shots, BRSW selected four as likelyassassination shots. One of these four,the third, was judged to have been firedfrom the grassy knoll; BRSW stated thatthe probability that this set of impulsesrepresents a shot from the grassy knoll is50 percent.

Judging that the principal limitation ontheir ability to make a more definitiveclaim was the microphone spacings forthe test shots, which led to the +6-millisecond window, BRSW suggestedthat WA be asked to try a theoreticalacoustic modeling technique. This pro-cedure was applied only to the hypo-thetical knoll shot. Several test shotswere examined, with a shooter locationon the knoll, and prominent echoes wererelated to Dealey Plaza objects. A theo-retical model of sound propagation inDealey Plaza, incorporating possiblevariations in shooter position, micro-phone position and velocity, and air tem-perature, was used to predict the relativetimings of echoes that would be expectedin the channel I recording if the segmentin question contained the sounds of agunshot fired from the grassy knoll. Ineffect, for every choice of shooter posi-tion, microphone position, microphonevelocity, and air temperature, WA coulddetermine the time of impulses of a hy-

SCIENCE, VOL. 218

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

pothetical test shot; using the binarycorrelation measure, they could corre-late these impulses with the channel Isegment identified by BRSW as a possi-ble knoll shot-with the significant dif-ference that matching pulses were re-quired to be within 1 millisecond ratherthan 6 milliseconds. WA did not try allpossibilities, but used the results of eachtrial to guide a search until they found ashooter position, and so on, for whichthe greatest number of predicted im-pulses in the theoretically reconstructedtest shot agreed with actual impulses onthe channel I tape. The best agreementwas found for a part of channel I that is0.2 second earlier than that suggested byBRSW. On seeing the WA results, theBRSW investigators agreed that the WAidentification should be used.

Evaluation of the BRSW and WA

Methodologies and Conclusions

A reliable analysis of the channel Irecording presents serious difficulties.The noise level is high, there is conflict-ing evidence about the location of theopen microphone, some of the back-ground sounds are difficult to interpret,the absence of certain expected back-ground sounds is difficult to understand,and the transmitting and recording sys-tems distorted the acoustic signals. Aspointed out by WA, to the ear the soundsresemble static much more than they dogunshots, and it is only the poor fidelityof the radio dispatching system thatmight permit the latter interpretation.Such static-like sounds could be generat-ed by a number of other acoustic, electri-cal, or mechanical sources in the envi-ronment and in the radio transmission,receiving, and recording equipment.Tests and analyses more discerning thanthe human ear are required to determinethe probable cause of the sound im-pulses. The WA analysis is ingenious butit is novel in some aspects, and both theBRSW and the WA echo techniques forgunshot location had not been appliedpreviously by either group to a situationwith as high a level of noise and distor-tion as this one.

Further, the BRSW and WA studieswere seriously limited by funds and fixeddeadlines. A number of essential tests toconfirm the analysis procedure and theinterpre.tations were omitted. The WAstudies, for example, were limited to thesingle conjectured grassy knoll shot. Theresults of such an analysis should not beconsidered reliable until the method hasbeen adequately tested on some othercases. In particular, the impulses conjec-8 OCTOBER 1982

tured to be sounds of gunshots from theTexas School Book Depository shouldhave been analyzed by the same method.Not only would this have provided acontrol on the method, it would alsohave provided a much stronger indica-tion of whether the open microphonewas or was not in Dealey Plaza at thecorrect time. Similarly, more of the testshots should have been analyzed to com-pare the observed echo patterns withthose predicted from structures identi-fied in the echo patterns with a differentneighboring microphone location.The original BRSW report claimed a

50 percent probability that there was anadditional shot from the grassy knoll.This claim was used as a justification forthe more detailed studies of WA. Theresult of WA's analytic echo predictiontechnique in the subsequent analysis ofBRSW's conjectured shot would appearto improve the credibility of the grassyknoll hypothesis. However, the commit-tee noted that the identification of shotsand impulses by BRSW was completelydifferent from that by WA, as demon-strated by the more than 200-millisecond(more than 200-foot) displacement be-tween the two identifications. Bargerpointed out that the two different identi-fications may be reconciled by assumingthat the BRSW echo pattern had beensubject to reflection from one additionalwall. Even with this interpretation thereremains a serious flaw, namely that theBRSW analysis missed the identificationthat WA considered to be the primaryone.The impulses selected for the BRSW

study were not always the largest ones.Frequently, large impulses were omittedand some impulses close to the noiselevel were retained. Far more impulsesdo not fall into the BRSW classificationof "probable sounds of gunfire" than do.Since the results of correlation coeffi-cient calculations are highly dependenton the impulse and echo selection pro-cess, it is especially critical that thescheme used to distinguish these soundsstand up to close scrutiny, with the pro-cess used being spelled out in detail sothat others can duplicate the analysis.From the published reports, it is impossi-ble to do so. Furthermore, weak spikeson the Dictabelt often are selected tocorrespond to strong patterns in the testpatterns, and vice versa.The conclusions of the BRSW analysis

were supported by some later interpreta-tions of photographic evidence as beingconsistent with a motorcycle in the pro-cession at approximately the positionindicated by their analysis. However, itis not certain that this was the motorcy-

cle with the open microphone, that itsradio was improperly tuned to channel I,that the open microphone was in DealeyPlaza, or that the relative times of thefour sets of impulses studied by BRSWand WA were consistent with the threeknown actual shots. There is evidence tothe contrary on all four of these pointsthat should not be ignored.No siren sounds are heard on channel

I at a time when they should have beenheard by an open microphone in themotorcade; sirens are not heard for ap-proximately 2 minutes after the impulsesattributed by BRSW and WA to assassi-nation shots, after which clear and un-ambiguous sounds from a group of sirensoccur on channel I. The sirens seem tocome from a group of at least threevehicles and the intensity of the soundfirst increases and then decreases. Thisis consistent with sirens heard at a sta-tionary point if the presidential motor-cade had passed close by. It is not thesiren sound expected if a motorcyclewith a stuck button was part of themotorcade. In the first quarter-mile ofthe trip to the hospital the presidentialmotorcade encountered a complex pat-tern of underpasses, roads, and ramps,but there is no trace of a siren sound inchannel I during this interval of time.This initial long absence of siren sounds,followed by the sounds of several sirenspassing by, suggests that the radio trans-mitter with the stuck button was not partof the presidential motorcade, but mayhave been on a parked motorcycle.James Bowles, police communicationssupervisor at the time of the assassina-tion, suggests that it was on a motorcycleparked at the police command post nearthe Trade Mart, where it would be natu-ral to have adjacent police radios tunedto different channels.The BRSW report concludes: "The

probability of obtaining just one matchby chance in any of 180 independent triesis equal to 5.3 x 10-2, or about 5%.Therefore, the probability that they ob-tained their match because the twomatched patterns were due to the samesource (gunfire from the knoll) is about95%." The WA report concludes with asimilar statement. Such statements donot allow for the existence of hypothesesalternative to the two primarily consid-ered (the hypothesis of gunshots and thehypothesis of impulses randomly locatedaccording to a Poisson distribution inrelevant sections of the Dictabelt). Vari-ous reasonable alternative hypothesesinclude nonwhite (non-Poisson) noise orother typical noise and static distribu-tions which are ordinarily lumped to-gether in time and therefore may give a

129

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

higher correlation with the nonrandomdistributions of test shot echoes. Fur-thermore, even if the only alternative toimpulses from a gunshot were the hy-pothesis of randomly located impulses, asingle observed result whose P valueunder the random location hypothesis is5 percent does not imply a 95 percentprobability that there was gunfire fromthe knoll (the P value or significancelevel in current statistical theory is theprobability, assuming the hypothesis tobe true, of observing data as or moreextreme than what actually is observed).The situation is analogous to that in acard game where the probability of thedealer receiving three aces is P = .44,but three aces going to the dealer on thefirst deal does not by itself indicate a 95.6percent probability that the dealer isdishonest if there is no prior reason forsuspecting him of cheating (2, p. 38).

In addition, the BRSW-WA calcula-tion of the P value for the hypothesis ofrandom pulse location is incorrect.There are several errors, of which themost serious is the failure to allow in theprobability calculations for the fact thatthe location of the shooter in the WAanalysis was adjusted to maximize thenumber of coincidences. With these cor-rections and a conservative adjustment,a P value as high as .223 can be obtainedfor the hypothesis of random location;this is much less impressive than theBRSW-WA value of .05 (2, p. 35). Fur-thermore, even if it were granted that thehypothesis of randomly located impulseson relevant portions of the tape were in

serious doubt, it would not follow thatthe alternative of gunfire from the grassyknoll was convincing. All plausible alter-natives to both of these hypotheseswould have to be eliminated, and noconvincing effort has been made in thisdirection.

After this study of the BRSW and WAreports, no member of the committeewas convinced by the arguments giventhat there was a grassy knoll shot. Themembers of the committee reached theirinitial negative conclusion prior to theavailability of the event timings and thesound spectrograms discussed below.

Timing Evidence from

Matching Features

A private citizen, Steve Barber ofMansfield, Ohio, wrote to the committeethat he was convinced from his ownlistening that there are clear instances inwhich phrases recorded on channel IItape were distinctly audible on the chan-nel I tape as well. This is naturally ex-plained by assuming that the motorcyclewith the open microphone (channel I)was near another police radio receiving atransmission from channel II, so thattransmissions over channel II would bepicked up by the open microphone andrebroadcast on channel I. In addition,there are simultaneous broadcasts by thedispatcher on channels I and II. Bothkinds of cross talk are perfectly clear inmany cases and would allow precise timesynchronizations between specific por-

tions of the two recordings. The timesynchronizations would not apply to therecordings in their entirety, becausechannel I ran continuously during theperiod of interest while channel II wassound-activated and operated intermit-tently.Four such matching sections on the

two tapes are quite clear, but they occurseveral minutes after the assassinationand involve various police communica-tions connected with the follow-up to theshooting. (They also provide a cleardemonstration of channel I heterodynessuppressing the recording onto channel Iof cross talk from channel II; we latershow that this suppression also occurs inthe interval containing the impulses andshows that the cross talk was recordedthrough a radio receiver.) Two eventsare especially important for fixing thetime of the section of tape analyzed byBRSW and WA. The first is a 4-secondfragment of speech that overlaps theconjectured third and fourth BRSWshots on channel I. Barber there identi-fies a phrase, which he says begins withthe words "hold everthing," as identicalto the phrase "hold everything secureuntil the homicide and other investiga-tors can get there," clearly recorded onchannel II. The significance of this pro-posed match is that the section on chan-nel I is concurrent with the last two ofthe conjectured BRSW shots, whereason channel II it is part of a clear se-quence of emergency communicationsthat followed the shooting and occurredapproximately 1 minute after the assassi-

Fig. 1. Composite photograph of sound spectrograms on channel I (top) and channel II (bottom). The audible "hold everything" phrase begins atapproximately zero on both channels, but there is no special significance to the exact location of zero time on either channel. The impulsesinitially identified by BRSW as arising from their conjectured grassy knoll shot occur above the arrow marked 145.15s, and those identified byWA occur 0.2 second earlier; the proper location for this arrow was determined by comparing this sound spectrogram with that of figure 5 in theBRSW report. The letters and black dots designate corresponding characteristic features. For reasons discussed in (2, p. 41), one recording wasspeeded up by 6.7 percent.

130 SCIENCE, VOL. 218

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

nation. It is, in fact, part of Sheriff Deck-er's instructions to his men in responseto the assassination. This time synchro-nization, if correct, would prove that theBRSW-WA conjectured shots were un-related to the sounds of the assassinationgunshots. The section of channel I re-cording with the BRSW-WA conjecturedshots would then correspond to a periodof time well after the assassination.The second important event is the

transmission "You want me . . . Stem-mons," which occurs several minutesafter the assassination and is intelligibleon both channels. It provides a commonreference point for timing events on thetwo channels. We used it to determinewhether the section of the recording con-taining the conjectured shots occurredbefore or after Chief Curry instructed themotorcade to "Go to the hospital."

Sound Spectrograms

Initially, the poor quality of the "holdeverything" portion of the recordingmade it appear unlikely that a convincinginterpretation of the badly garbledspeech on channel I could be made, andthe committee was aware that a listenerto a garbled message will often be con-vinced that he has heard what he hasbeen coached to hear.For these reasons, arrangements were

made for members of the committee touse the excellent sound analysis equip-ment of the Federal Bureau of Investiga-tion's Technical Services Division to ob-tain sound spectrograms ("voiceprints")of the relevant communications on chan-nels I and II. The spectrograms wereprepared under the supervision of com-mittee members. The sound spectro-grams first reproduced were from taperecordings provided by the Dallas police,but a sound spectrogram with a similarpattern for the "hold everything" phraseon channel I was also made from a tapesupplied by James Barger, essentiallyidentical to that used in the BRSW analy-sis; later sound spectrograms were alsomade from new high-quality magnetictape copies of the original channel IDictabelt and channel II Audographdisk. A sound spectrogram is a plot withelapsed time along the horizontal axis,frequency along the vertical axis, and thedarkness of a trace representing the in-tensity at the frequency of the trace.We then made spectrograms of the

"hold everything" sections. Figure 1 is aphotograph of composite sound spectro-grams for the full 4-second message. Thebeginning of the "hold everything"phrase is approximately at zero on the

8 OCTOBER 1982

4

0

0 1 2 3T" (sec)

Fig. 2. Plot of T' against T" for correspondingcharacteristics. The linearity of the curve

demonstrates the validity of the identification.The straight line is a plot of the equationT' = -0.0216 + 1.0593 T", which in turn is a

robust linear regression fit to the plottedpoints. The analysis leading to this figure isgiven in (2, p. 41). The point furthest off theline is at T" = 1.195 seconds and is for theincorrectly identified characteristic I, as dis-cussed in (2).

time scales. The impulses for the conjec-tured grassy knoll shot begin approxi-

mately at the arrow marked 145.15s (thetime of the conjectured grassy knoll shoton the BRSW time scale) and the WAimpulses occur 0.2 second earlier. Theblack dots mark 27 corresponding fea-tures on the two channels.

It is apparent from Fig. 1 that there is a

correlation between parts of the soundspectrograms of the two channels, even

though the channel I recording has muchmore noise. The correlation is muchmore impressive when the spectrogramsof the two channels are compared indetail. It is particularly striking when one

realizes that only the initial second of the"hold everything" phrase can be heardclearly on channel I, yet the sound spec-

trograms contain numerous matchingfeatures for the entire 3.5-second se-

quence (note, for example, the match inthe final segment from 3.2 to 3.6 sec-

onds). In all cases of matching features,the text of the messages and the signalintensities show that a signal from chan-nel II was duplicated on channel I andnot the reverse.

The sound spectrograms present muchmore convincing evidence in this case

than in their application to speaker iden-tification. In the latter case, words spo-ken at different times, supposedly by thesame speaker, are compared, and a

trained interpreter is often required toexplain why the subjective match is sig-nificant. In the present case, the need isto identify two identical messages ex-

tending over a 3.5-second interval. Notonly must individual parts of the twosound spectra be alike, but they mustoccur at exactly correct time intervalsand with exactly matching frequencies.The existence of these required time andfrequency correlations between the twochannels imposes rigid constraints on themessages to be matched. Furthermore,all sounds that appear on both channelsare useful in correlating the channels,even though some are not spoken words.For example, in listening to channel II itis apparent that there is an intermittenttone that contributes to the flat portionscommon to channels I and II. However,this tone varies in both amplitude andfrequency and is also useful in correlat-ing the two channels.

Analysis of Sound Spectrograms of

"Hold Everything"

The committee used three methods inaddition to visual inspection to deter-mine whether the sound spectrograms ofchannels I and II contained signals fromthe same source (2, p. 41). In the firstmethod we identified characteristic fea-tures that were present on both spectro-grams and then determined the relationbetween the times of occurrence of thetwo sets of features. Twenty-seven fea-tures were selected (2, p. 43). The exis-tence of correlations between the twospectrograms over a long time intervalcan be demonstrated by plotting T', thetime coordinate of the channel I spectro-gram, as a function of T", the time co-ordinate of the corresponding charac-teristic on channel II. The results areshown in Fig. 2. The linearity of the plotshows that the similar characteristics ofthe sound spectrograms of the two chan-nels follow the same time sequence, asthey must for one to be cross talk fromthe other. A linear fit to the recordedpoints gives the equation in Fig. 2; theslope of the line or ratio of recordingspeeds is 1.059 ± 0.002, which corre-sponds to a 5.9 + 0.2 percent net differ-ence in the recording speed. The ratio ofrecording speeds independently inferredfrom the measured frequency ratios ofthe same points is 1.064 + 0.006. Theprobability of obtaining such close agree-ment by random occurrence of the fea-tures at their observed average spacingwould be about 2.1 x i0-13, and theprobability of randomly obtaining suchgood agreement on the frequency ratio ofthe points is about 2 x 10-10.The second method is the calculation

of the effective relative speeds from thefrequency ratios for five sections with

131

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

particularly well-defined frequencies onthe two channels. Such a calculationgives a ratio of recorder speeds of1.062 ± 0.005 in excellent agreementwith the value in the precedinig para-graph.To help in the visual recognition of

0.5

0.4

0.3

0.2

0.1 j

0,^/ 'IW\' 9' 's?''

4VtN V

A'I I I.N,& IW A, ;P, I

i '; "^

similarities of the two patternspectrograms have been madespeed of channel I effectively ch;6.7 percent. The results are sFig. 1. Both frequencies and timtwo channels now appear to beble.

I'.,

,V,ry4.

IV .-

0 10o 200 300 400 500 600Fig. 3. Cross correlation between "hold everything" segments of channel I and clsound spectrograms with time scale slightly compressed to produce the best correlatiThe curve is produced by sliding 2.50 seconds of channel I along 10.00 seconds of ci0.01 second at a time, using frequencies in the band 600 to 3500 hertz.

0.8

0.6

0.4

0.2

0

.III

I.I111II

.

,i\ A :~~~~~~~~~'i\

0 100 200 300 400 500 600

Fig. 4. Cross correlation between "You want me ... Stemmons" segments of chanichannel 11 sound spectrograms with time scale slightly compressed to producecorrelation peak. The curve is produced by sliding 2.50 seconds of channel I aloseconds of channel II, 0.01 second at a time.

132

s, sound In the third approach, the channel Iwith the and channel II recordings were digitizedanged by and the short-term acoustic spectra were,hown in taken and stored in a digital computer.Les of the The printouts of these spectra are similarcompati- to Fig. 1. These digital spectrograms

were computed directly from magnetictapes and did not involve use of the FBIsound spectrogram equipment. Many ofthe features observable in the analogspectrograms can be seen in the digitizedspectra, but no use was actually made ofthe spectrogram patterns. Instead, theactual data were used to test certainhypotheses, without human interven-tion. An objective measure of similarityof two spectral matches is obtained fromthe cross-correlation coefficient (2, p.57), which would be reduced if one of therecordings were played at the wrongspeed, or if the recording at one timewere compared with the same or a differ-ent recording at a different time.The first cross-correlation coefficients

t were made from the channel I and chan-nel II recorded copies that were used inpreparing Fig. 1. It was found that thebiggest peak for the cross-correlationcoefficient occurred for a relative warp(speed ratio) of 1.06, in agreement withthe values estimated by the other twomanual approaches; a 1 percent devi-

700 ation of warp from optimum diminishedhannel II the peak substantially. However, thation peak. channel II copy contains many repeats,iannel 1, caused by the Gray Audograph machine

in playback. Accordingly, another tapecopy was prepared by members of thecommittee directly from the original Au-dograph plastic disk and by use of astandard turntable and tone arm, withoutcompensation for the fact that the diskwas originally recorded at constant lin-ear track speed. Figure 3 shows thecross-correlation coefficient for the"hold everything" segments when therelative speed was selected to give thelargest peak and the 750 correlation coef-ficients were obtained by sliding 2.50seconds of channel I along 10.00 secondsof channel II, 0.01 second at a time,using frequencies in the band 600 to 3500hertz. For comparison, the cross-corre-lation coefficients of the unambiguoussegment "You want me ... Stemmons"

J,, are plotted in Fig. 4. The shape of the'P>zt peak is very similar to that for the "hold

everything" segment. The background issomewhat smoother because there is lessnoise in channel I at this time. Channel I,however, in neither case gives a perfect

700 reproduction of channel II. It has lostelI ad some of the high and low frequencies,

the best and as one would expect there are tonesong 10.00 present on channel I that are not on

channel II.

SCIENCE, VOL. 218

L--

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

The narrow peaks of the cross-correla-tion curves show by an objective testthat the "hold everything" segment ofchannel II is present on channel I at thesame location as the acoustic impulses.There is no doubt that the voice (andother) sounds of channel II are presenton channel I to an accuracy in locationcorresponding to a few milliseconds.We consider three sets of results to

constitute overwhelming evidence thatthe "hold everything" sections of thetwo recordings are traceable back to asingle acoustic signal from channel II. Ifthere is no overrecording on channel I(as we show below), the correspondencebetween these two recordings of "holdeverything" would be conclusive evi-dence that the events analyzed byBRSW and WA were not the assassina-tion shots, since we know from channelII that the "hold everything" transmis-sion was made at least 50 seconds afterthe chief instructed the motorcade to"Go to the hospital."

Possibility of Superposed Recordings

The committee considered seriouslythe possibility that the impulses analyzedby BRSW and WA were overlaid at alater time by the "hold everything" mes-sage. Such an overrecording could haveoccurred if the Dictabelt or the recordinghead was knocked backward by about 1minute in the first minute after the assas-sination, or if a' new Dictabelt copy,made by audio coupling while a channelII recording was playing in the back-ground, was substituted for the original.The committee concluded that this wasnot the case on the basis of (i) physicalexamination of the Dictabelt for indica-tions of overrecording or of substitutionof a copy for the original; (ii) the unlikelynature of any of the highly contrivedscenarios required to provide such anundetectable overrecording either acci-dentally or deliberately, (iii) the compati-bility of the timing implied by the "hold

everything" identification with otherfirmly established evidence, and (iv) theconclusive acoustic evidence on the Dic-tabelt itself that the cross talk recordingswere made through a radio receiver withautomatic gain control (AGC). These dif-ferent forms of evidence are all compati-ble with the recordings being made at thesame time, and some are incompatiblewith the hypothesis of later superposedrecordings by audio or direct electricalcoupling. Only the evidence of category(iv) will be reviewed here.The digital analyses of the sound spec-

tra can be used to demonstrate that thechannel II imprint on the channel I re-cording was already present at the chan-nel I receiver and was not added later inthe recorder or as an overrecording. Theby-radio nature of channel II cross talk isdemonstrated by its detailed behavior inthe presence of channel I heterodyneswhen another channel I transmitter iskeyed on with a more powerful carriersignal. The frequency offset between thetwo carriers gives rise to a heterodynetone in the channel I recording. Howev-er, the channel I receiver was fitted AGCto hold the output level aproximatelyconstant; as a result, the cross talk sig-nals decrease in intensity in a few tens ofmilliseconds (as does any residual trans-mission from the original open micro-phone). At the end of the channel Iheterodyne, the AGC gradually in-creases the receiver gain, and signals onthe open-microphone transmission in-crease in intensity in the recording. Anexcellent probing signal for the channel Igain would be a channel II steady toneacoustically coupled from the field loud-speaker to the open-microphone trans-mitter. This would come in at a constantlevel, and the variation in level on thechannel I recorder should mimic theAGC action if the channel II signals werepresent in this way. Inspection of thedigital spectrogram (and digital tabula-tions of the data) shows that numerouschannel II brief tones have a constantlevel from beginning to end. A crucial

demonstration is provided by the chan-nel I heterodyne beginning at time 32.02seconds in one of the spectrograms. Theunderlying channel II brief tone is sub-stantially reduced in intensity at the be-ginning of the channel I heterodyne andgradually grows back when the channelII brief tone results after the channel Iheterodyne ceases [figures B-7 and B-8in (2)]. This behavior is validated bysimilar channel II brief tones underlyingchannel I heterodyne signals in the "Youwant me . . . Stemmons" phrase and in aphrase "I'll check" that is also presenton both channels.

Conclusions

For the reasons discussed above andin its formal report (2), the committee onBallistic Acoustics unanimously con-cluded that the acoustic impulses attrib-uted to gunshots were recorded about 1minute after the President had been shotand the motorcade had been instructedto go to the hospital, and that reliableacoustic data do not support a conclu-sion that there was a second gunman.The committee's charter asked it to

recommend the tests, analyses, and eval-uations needed to obtain better informa-tion from the recordings. If there were tobe further studies of the channel I re-cording in the hope of demonstrating thevalidity of the conjectured shot from thegrassy knoll, the information listed in thecommittee's report (2, p. 92) should besought. However, the evidence againstthe conjectured grassy knoll assassina-tion shot is already so strong that webelieve the results to be expected fromsuch studies would not justify their cost.

References and Notes

1. Appendix to Hearings Before the Select Com-mittee on Assassinations of the House ofRepre-sentatives Ninety-Fifth Congress (GovernmentPrinting Office, Washington, D.C., 1979), vol. 8.

2. Commission on Physical Sciences, Mathemat-ics, and Resources, National Research Council,Report of the Committee on Ballistic Acoustics(National Academy Press, Washington, D.C.,1982).

8 OCTOBER 1982 133

on

May

22,

201

3w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from