The Spy Files, Wikileaks. España vol I

8/3/2019 The Spy Files, Wikileaks. Espaa vol I

1/33

!ro$uc's) +vervie/

Currently deployed fieldproven Biometric SpeakerIdentification andVerification Technology.


2/33

!"#$%&'()**+(

,-.&/

! 0!,12(3,!4256,17

! 3850!9

! 87,7

! 37:

! 1!;1


3/33

!"#$%&'()**+(A12ITI+ !ro$uc')s +vervie/

0$>"&(3>$?'>"@A(B&C(5&"D-$E$FC(G$'(7&"H'>#C(

! Use of voice evidence in court.

! Identification of unknown intercepted voices

! Intelligence and crime prevention using phone conversationsmonitoring.

! Protection against Identity fraud

0$>"&(3>$?'>"@


4/33


IDJ#(>@(0$>"&(3>$?'>"@K

0$>"&(3>$?'>"@

Extraction of unique characteristics of vocal tract,from voice sound waves.


5/33

A12ITI+ !ro$uc')s +vervie/

IDJ#(>@(>-(#D&(L$>"&K

! ,-G$'?J#>$-(EJC&'A

! What the person is saying. The content of the speech.

! Can be automatically obtained using classical speech recognition technologies.

! 3&DJL>$'JE(EJC&'A

! How do you say itA Your accent, your emotions, your culture, your country DlanguageE,

!"#$%&'!#(!)

! This is the level used by impersonators, and the level the brain mainly uses to identify

a person Dand obtain additional informationE

0$>"&(3>$?'>"@

5M2(3,!4256,1(


6/33

M$O(0$>"&(3>$?'>"@(O$'P@K

! 5oice mo$el 8aka as ;voice

! Using a minimum of M seconds of net speech, we compare the utterance with the model

*'+$&,!*-'$*$./%&0#12$3"#$"-4"#0$!"#$/%&0#$!"#$5&0#$60&,*,7#$-!$-/$!"*!$!"#$8'9'&:'$;&-%#$

belongs to the same person.

! 6&@HE#@

! epending on the application, different techniques are used to .'&05*7-$?'>"@


7/33


0$>"&(3>$?'>"@A(B&C(1$?Q>#>L&(8.LJ-#JF&@

Driver& Increasing internationa0terroris1 and cri1e using di44erentt56es o4 6hones8 the Internet8 and in

1u0ti60e 0anguages9

Voice is the only biometric characteristicobtained remotely

It is already available in abundance.

Includes valuable information besidesbiometric information

asy to integrate in existinginfrastructure: Telephone networks, VoIP,etc..

0$>"&(3>$?'>"@


8/33

Agni'io)s CniDue A$?'>"@


9/33

8F->#>$A(R>$-&&'>-F(R'$.H"#@

G$'(0$>"&(M"J#>$-@


Present VoiceEvidence in court

Identification ofunknown voice during

police investigation

Speaker spotting tofilter mass phone

interceptions.

0$>"&(3>$?'>"@


10/33

ST(U$'&-@>"(2L>.&-"&(>-(1$H'#(M&J'>-F@(


Forensic Labs relying only on Phonetic and

Linguistic techniques have limitations:Impossible to handle unknown languages,

difficult to handle great number of calls,!"#$

Intercepted phoneconversations are used asevidence in court. Experts

are required to performSpeaker Verification

MARKET NEED

SOLUTION

Tool for biometric voice verificationdesigned for Forensic Experts. It

provides the precision andreliability required in court hearings

PROBLEM

3850!9


11/33

7Q&JP&'(,V(>-(U$'&-@>"@

!1EJ@@>"JE(8QQ'$J"DA(-FH>@#>"WQD$->"@

! An expert in languages and phonetics compares both voice samples(Suspect and unknown)

!7&?>JH#$?J#>"

! Using computer voice analysis programs, sonograms, formants and othertechnical information are compared.

!0$>"&(3>$?'>"@

! Computer generates voice print, unique to speakers physical vocal tract.

! Computer compares suspect and unknown voice samples and provides amatching score.


3850!9


12/33

A$$e$ 5alue) of 5oice Giome'rics in Horensics

! 7Q&JP&'(,V(L&'>G>"J#>$-(H@>-F(QDC@>"JE("DJ'J"#&'>@#>"@($G(#D&(

L$"JE(#'J"#X(>-@#&J.($G(#D&(OJC(#D&(Q&'@$-(@Q&JP@(

! >&%*7$!0*%!$.7#*;#/1$8'-?8#$@#*!80#/$-'$;&-%#$/&8'+$:*;#/A

! 0$>"&(3>$?'>"@(J..@(-&O(J-.('&E&LJ-#(>-G$'?J#>$-(#$("EJ@@>"JE(#&"D->YH&@T

! epending on the case DLanguage, unique phonetic features, ..E it is more or less

relevant.


3850!9


13/33


3850!9A(R>$-&&'>-F(R'$.H"#(G$'(U$'&-@>"(&@

! 2-#'C(R'$.H"#

! Widely used in urope, Latin merica

and sia

!First working system in SpanishPolice DIdentivox 2000E

! xpert reports presented in many

courts globally

V'T(BZ-[&E(\=->LT(4J'%H'FX(]&'?J-C^(G$'?&'EC(#D&(P&C(&/Q&'#($G(#D&(

]&'?J-(U&.&'JE(R$E>"&T(!-&($G(#D&(?$@#('&@Q&"#&.(@Q&">JE>@#@(O$'E._O>.&T

!"#$%&'#('')#*+,)-#.$'#/01234#%*.56%.,7#+8+.'6#999,)#65:'#.$%)#;


14/33


BATVOX reduces learning curve, guiding the new userwith a wizard, qualifying results and helping

interpretation of data

3850!9


15/33


"BTVO provides the

expert Likelihood Rate DLREresults

"BC$-/$!"#$D/!*!#-of-the-*0!E$metric to measure evidencestrength in court hearings

3850!9


16/33

7H""&@@(7#$'CA(1$?%J#>-F(1$''HQ#>$-

! R'$?>-&-#(EJOC&'(>-(4JEJC@>J(OJ@(#JQ&.(%'$P&'>-F(#D&(JQQ$>-#?&-#(($G(#$Q(`H.F&@

!0>.&$(Q$@#&.($-(N$H5H%&("'&J#&.(J(#H'?$>E(>-(4JEJC@>J-(Q$E>#>"@


%&'&()*&+,-+"*,./0012"*/+,-3!+#(,1)!4,-567879:),20/41#"),&+4,,,,,,,,,,support to verify the voice identity, and present evidence in court

3850!9


17/33

7H""&@@(7#$'CA(((((((((((((((((((((((((((((((((((((((((((1$?%J#>-F(,-#&'-J#>$-JE(!'FJ->[&.(1'>?&

! 8E%J->J-(]J-F(J"#>-F(>-(]&'?J-C("$??>##&.(>.&-#>#C(G'JH.(#D'$HFD(#D&(#&E&QD$-&T

! ffecting more than Z00 German elder citi\ens

! ]&'?J-(Q$E>"&('&"$'.&.(?$'&(#DJ-(:X***(QD$-&("$-L&'@J#>$-@(%$#D(J?$-F(#D&(FJ-F(?&?%&'@(#D&?@&EL&@X(J-.($G(FJ-F(?&?%&'@(@O>-.E>-F(#D&>'(L>"#>?@T

! Conversations were in German, nglish, lbanian and Sinto. Some ofthe gang members spoke three different languages in various tapedconversations.


German Forensic Experts verified the identity of the gang members*+;/';!4,*+,"


18/33

IDJ#a@(-&/#KA(SA(;(,.&-#>G>"J#>$-(1JQJ%>E>#>&@


1:1 Forensicevidenceverifications

1:N for Lawenforcementidentifications

87,7


19/33

)T(?>-JE(,.&-#>G>"J#>$-(H@>-F(0$>"&


Police have a large amount of voicescaptured but identification was not possible.

Traditional methods not suitable for largevoice database searches. Sometimes voiceis the only available personal characteristic

of a criminal

Police obtains voice fromunknown person involved in

crime. (Bomb threat, lawfulphone interception, terroristvideo). Identification is

needed

MARKET NEED

SOLUTION

Tool for unknown speakeridentification using voice

biometrics. Automatic biometricsearch easy to perform with

basic training

PROBLEM

87,7


20/33

87,7A(0$>"&(3>$?'>"(VJ#J%J@&(G$'(


21/33

87,7A(4$.HEJ'(J-.(@"JEJ%E&


87,7

DB Manager

GUI

Browser 1

Engine 1

DB

Task Manager

Application

Server

Presentation Tier

GUI

Browser 2

GUI

Browser N

Business Logic

Logic Tier

Data Tier

Engine Manager

Engine 2

Engine N

Queue


22/33

87,7A(4$.HEJ'(J-.(@"JEJ%E&


Browser 1

Browser 2

Browser 3

Engine 1

Engine 2

Engine NDB Server

Internet

Application Server Task Manager Engine Manager

ASIS MULTICORE

Active MQ


23/33


From 1,000 up to 100,000 voice models per minute.

SP

From ^0,000 up to 200,000 voices stored DBithout6er4or1ance degradationE

From ^ up to 1,000 simultaneous connections.

aximum capacity

Optimi\ed for HP and Linux Suse ngines: Proliant L320 with eon Quad cores Task anager: Integrity 300 with Itanium ual

core.

Hardware/OS

87,7


24/33

87,7(R&'G$'?J-"&A(141


AGNITIO s forensic bench test Database.

How to measure

performance on 1:NsystemsA

CC: Cumulative atchingCharacteristic

Probability of your targetbeing among the first Ncandidates

87,7


25/33

us'omers are S'ar'ing 'o Tes' an$ 'o Cse i'


A Latin A1erican securit5organiation9 Ehe 0argest :Fnvoice test 6er4or1ed unti0now9 A01ost :G8=== voicesand 1ore than :9HM tria0s

!"

!#

$"

$#

1""

1 11 21 '1 (1 #1 )1 *1 !1 $1

1'+("" ,-./0

1#"3/3

4-516!17

4-5#6$#7

4-51"6$*7

4-589:3;/093 ?>9093@

4-51 *27

4-5' $'7

4-5# $)7

An uro6ean 6o0iceorganiation9 Eest with rea04orensic audios9 S1a00erdata@ase9

87,7


26/33


3850!9 87,7

R'&">@>$- aximumaccuracy in a1:1 verification

Obtain a list ofcandidate. Theverification can then berefined in a forensiclaboratory

2''$'@ inimi\e falsepositives inimi\e false negatives

=@J%>E>#C( For experts withan importantbackground inforensics

For operatives with nobackground in audio andbiometrics

7Q&&.( Not critical Critical: thousands ofidentifications perminute

87,7


27/33

87,7A(0$>"&(,.&-#>G>"J#>$-

!


28/33

:T(,;52


29/33

37:A(B&C(G>E#&'>-F(#$$E(G$'(,-#&EE>F&-"&(8F&-">&@

! 7R28B26(V25215,!;(,;(4877(52


30/33


37:


31/33


BTVO

SIS BS3

FISCOISIBISCentral

Criminal ata

OtherIdentification

atabases

Selected udios

atacrunching

Key wordspotting

BS3

37:


32/33


3850!9 87,7 37:

R'&">@>$- aximumaccuracy in a1:1verification

Obtain a list ofcandidate. Theverification canthen be refined in aforensic laboratory

Obtain a list ofcandidates callswhere the target isspeaking.

2''$'@ inimi\e falsepositives

inimi\e falsenegatives

inimi\e falsenegatives

=@J%>E>#C( For expertswith animportantbackground inforensics

For operatives withno background inaudio andbiometrics

For Intelligencegencies doingcriminal prevention.

7Q&&.( Not critical Critical: thousandsof identificationsper minute

Critical: Several callsper minute.


33/33

5DJ-P@(G$'(

C$H'(J##&-#>$-T

U$'(GH'#D&'(>-G$'?J#>$-X(QE&J@&("$-#J"#A

milio artne\ CO

emartine\agnitio.es

mmanuel erome VP HLS Business evelopment

ederomeagnitio.es

Christian oreno HLS Business evelopment

chmorenoagnitio.es

Beatri\ Gon\le\ Sigen\a HLS Business evelopment

bgon\ale\agnitio.es

!"#$%&'()**+(

Documents

The Spy Files, Wikileaks. España vol I