The image of Saddam Hussein appeared on Iraqi television Monday for the second time since last Wednesday's strike on a bunker where he was believed to be hiding. He looked far more vigorous than he did in his TV speech Thursday. But once again, intelligence officials wondered whether the face on television was Saddam's or that of a double, standing in for a dead or wounded leader. They suspect it was the real Saddam, although the tape may have been prerecorded. But they are turning to computer analysis of his speech to be sure.
After last Thursday's appearance, it was widely reported that computerized face-recognition technology had helped officials determine the identity of the televised visage. That's unlikely, say experts in the technology. Face-recognition systems are far from perfect at matching faces even under controlled conditions, let alone on a poor-quality video. Says Frances Zelazny of Identix, a company specializing in face-recognition technology: "The system returns a probability measurement ... but it won't tell you whether it was Saddam or not."
"Voiceprint" technology is far better developed. Relying on the unique characteristics of an individual's speech, it is a spinoff of commercial voice verification applications that have burgeoned in the past few years. Banking customers who want to conduct transactions on the phone can "enroll" their voiceprint in the bank's database by speaking their account numbers. When they phone in later and repeat the same number, the computer fetches the customer's original voiceprint from the database to compare it against the current speaker and verify the customer's identity.
Alas, Hussein is not likely to have recorded a voiceprint for any database. Instead, military analysts have fed every known scrap of recorded speech from him into a computer to produce an algorithm, or mathematical description, of how Hussein sounds. Intelligence officials used the resulting verification filter to determine that last Thursday's video probably was the real Saddam, and they are analyzing the new video in the same way.
Voiceprint analysis also helps them filter through tens of thousands of intercepted audio messages looking for Hussein's distinct voice characteristics. "That's the needle-in-the-haystack kind of search," says Larry Heck, director of speech research and development at Nuance, a voice verification company based in Menlo Park, Calif. "In a commercial system, the user claims their identity. In a reconnaissance system, it's scanning many channels of audio, asking, 'Is this Saddam Hussein, or not?'"
Whether the answer is reliable depends on how accurate the original voiceprint is and on the quality of the broadcast or intercepted message. Funding from the military and intelligence agencies has made the technology far more powerful over the past decade, Heck says. For example, researchers can now use sound patterns to identify the type of microphone the speaker is usingwhether a telephone handset or a cellphone hands-free systemand then cancel out the extraneous noise that kind of microphone typically generates.
But voiceprint analysis isn't infallible, even when running on a supercomputer. The real world throws too many variables into the process. Simple static or background noise can gum up the system. Reading from a prepared text rather than speaking extemporaneously also throws things off. Even a bad head cold or sore throat can complicate thingslet alone the aftereffects of a near miss with a 2,000-pound bomb.