Title: Alvin, Simon, and… Al Qaeda? Finding Modified Voice in SIGINT Traffic

Release Date: 2018-01-19

Document Date: 2010-06-02

Description: This February 2010 post from the internal NSA newsletter SIDToday draws attention to voice-masking technology used by subjects of surveillance to frustrate the agency’s voice-recognition efforts: see the Intercept article Finding Your Voice, 19 January 2018.

Document: DYNAMIC PAGE -- HIGHEST POSSIBLE CLASSIFICATION IS
TOP SECRET // SI / TK // REL TO USA AUS CAN GBR NZL

(S//SI//REL) Alvin, Simon, and... Al Qaeda? Finding Modified Voice in
SIGINT Traffic

FROM:

Researcher, Human Language Technology Research Office (R64)

Run Date: 06/02/2010

(TS//SI//REL) Voice modification isn't only employed by the
makers of Alvin and the Chipmunks; it is also being used by
SIGINT targets in Yemen, Afghanistan, Iraq, and elsewhere.

In the last several months, multiple targets used voice
modification technology to distort their voices, most likely to
avoid identification by intelligence agencies.

(TS//SI//REL) In late 2009, CT analysts found two audio cuts
associated with Al Qaeda in the Arabian Peninsula (AQAP) in
Yemen that contained voice that had clearly been modified
(the talker sounded like a character from Alvin and the
Chipmunks). R6 and CES (S31) were asked to help recover
the original voice so that the talkers could be identified.

Within two weeks, we had created a profile of the modification technique employed by the AQAP
handset. CES has deployed a capability (now available at "go saas") to undo the modification .

(TS//SI//REL) Next we asked the question: is it possible to automatically discover modified voice
in SIGINT traffic? If so, then this could become a powerful tool for target discovery. In the case
of AQAP, this is exactly the capability that was required: after the December 25 bombing
attempt in Detroit , many AQAP selectors in Yemen went quiet. Some of these targets had used
voice modification in late 2009; were they still employing voice modification but on different
phone numbers?

(TS//SI//REL) In response to this requirement, R64 developed a software package called HLT
Lite to automatically discover modified voice. HLT Lite makes use of three core human language
technology (HLT) algorithms: speech activity detection , anomaly detection , and automatic
speaker recognition . HLT Lite was deployed to the Special Collection Service (SCS) site in
Yemen on 9 March, 2010 to search for modified voice in JUGGERNAUT data. After being
operational for only 30 minutes, HLT Lite discovered a cut containing modified voice. Since then,
HLT Lite has scanned over 1,000,000 audio cuts and discovered at least 80 confirmed
examples of modified voice. Four phone numbers have been tasked and many others remain
to be analyzed.

(TS//SI//REL) Modified voice is not restricted to AQAP in Yemen. We used HLT Lite to search
500,000 NUCLEON cuts and discovered 76 examples of modified voice across several Product
Lines, including Afghanistan/Taliban, Sunni Extremism, Lashkar-e Tayyiba, and Southwest Asia
Narcotics. So if you hear SIGINT traffic that sounds distorted, it may not be garbled or corrupted
. . . it may actually be modified! You can listen to examples of modified voice here .

(TS//SI//REL) Does this article sound familiar? Long-time SID today readers will remember the
March 2007 article " Alert: Voice Masking Is Discovered in SIGINT ."

(S//SI//REL) Note: The above image is a spectrogram of modified voice from a cut associated
with AQAP in Yemen. Have comments on this topic? Post them on the SIDtoday Blog .

"(U//FOUO) SIDtoday articles may not be republished or reposted outside NSANet
without the consent of S0121 (DL sid comms)."

DYNAMIC PAGE -- HIGHEST POSSIBLE CLASSIFICATION IS
TOP SECRET // SI / TK // REL TO USA AUS CAN GBR NZL
DERIVED FROM: NSA/CSSM 1-52, DATED 08 JAN 2007 DECLASSIFY ON: 20320108

e-Highlighter

Click to send permalink to address bar, or right-click to copy permalink.

Un-highlight all Un-highlight selectionu Highlight selectionh