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

(TS//SI//REL) Alert: Voice Masking Is Discovered in SIGINT
FROM: Adolf Cusmariu, HLT Operational Technologies (S24211), and
 Deputy Chief, Levant Sunni Extremism (S2I24)
Run Date: 03/22/2007

(TS//SI//REL) Use of voice-masking devices complicate efforts to "voice ID" SIGINT targets.
(TS//SI//REL) Earlier this month, Counterterrorism (S2I) Levant Sunni Extremism language
analysts noticed that the intercepts of two prominent Al-Qai'da associated extremists in the
Levant contained voice masking; the audio either sounded far too high, or too low in pitch.
(TS//SI//REL) In one instance Saudi-based senior extremist financier Abu-Ubaydah used the
voice-masking tool throughout his conversation dealing with his financing the travel of Yemeni
suicide bombers to both Lebanon and Iraq. In another instance, Syria-based senior foreign
fighter and facilitator Abu-Ghadiyah began in masked mode, went unrecognized by the
recipient, turned the masking device off, and both had a good laugh.
(TS//SI//REL) Of course, this is no laughing matter, since all efforts to identify any speaker both human and automated - will fail when voice alteration is used. But why would a caller do
this at all? Surely the recipient would also fail to recognize the speaker, so absence of this very
human method of sound identification would only induce suspicion and mistrust. Disturbingly,
could some terrorists be aware of efforts by SIGINT agencies to identify them by their voices and
thus have begun to resort to masking?
(TS//SI//REL) The devices used to alter voices in the few cases intercepted so far may not have
been all that sophisticated. Simple changes in sampling rates done in mathematical software
made the digital audiofiles sound more normal; in fact, two specific intercepts may have
contained the same speaker. Still, automated detection of such anomalies will be difficult.
(TS//SI//REL) A way to undo masking for an intended recipient would be to use the same device
and agree on settings prior to each transmission, so that unmasking could be done near-real
time. For example, if the sampling rates were increased by 25% to make the speaker's pitch
sound far too low, sampling rate reduction by the same amount at the other end would return a
normal voice. Even simpler, if only voice messaging is used, the masked recorded audio could
be undone after playback through an (un)masking device.
(TS//SI//REL) Because speaker identification services will continue to provide a key piece of the
puzzle in tracking and capturing terrorists, voice-masking techniques need to be understood
early. Tools developed for other intelligence applications are available and will be acquired to
study this emerging problem.
(U//FOUO) Do you have any comments on this topic? Post them on the SID today 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