September 6, 1995

Researcher develops system to identify criminals by their voices

In theStar Trek television shows and movies, the computer on the Enterprise knows which crew member is speaking by voice alone. Current technology is not that sophisticated, but an N.C. State computer scientist is working to make it so.

Dr. Rick Klevans, a researcher in the Department of Computer Science at N.C. State University, is perfecting his automatic speaker recognition system for forensic use. His work, funded by the North Carolina State Bureau of Investigation, may one day allow law enforcement officials to use computers to identify criminals by their voice.

Joel Rosch, director of Research and Development for the SBI, said that the system Klevans is developing will be useful when bomb threats happen to be recorded. (At the time of the Oklahoma City bombing, a number of bomb threats were made by phone in North Carolina.)

Klevans' work is not the first attempt to identify people by their voices. In the 1960s, a group of researchers believed that voiceprints--a picture of what speech looks like--would be useful in identifying people by their voices. The research stalled after a controversy arose over inadequate testing, and scientists shied away from forensic uses, pursuing instead commercial applications such as secured entry systems.

The forensic application will address two areas: verification and identification. Speaker verification determines whether two voices are the same. Speaker identification matches an unknown speaker to an individual in a group of known speakers.

For example, a threatening call by a woman is recorded, and officers trace it to an office where only six women work. Klevans' system would compare the caller's voice to those six voices and identify the the best match among the suspects.

In some cases, such as a bomb threat called in from a pay phone, a database of known speakers will not be available. Potentially, anyone in the country could have made the call. Klevans believes that problem could be partially solved by making recordings of anyone who is arrested--just as fingerprints are taken--and building a huge database of known voices.

The incriminating voice can be obtained from different sources: phone taps, eavesdropping equipment, undercover tapes, 911 tapes, and even telephone answering machines. When brought in for questioning, the suspects will be recorded saying anything, and the original recording will be matched against these recordings to produce the most likely suspect.

To achieve that end, Klevans first identifies usable information in the speaker's voice such as the way vowels are pronounced. A computer extracts these features from the voice yielding a set of numbers that models what the speaker sounds like. The computer compares models of unknown speakers to known speakers to find a match.

The computer then determines a degree of confidence in the match by using statistical techniques. In one experiment, Klevans used his system to identify individuals from a database of 230 male speakers. On average, the system eliminated 98.8 percent of the population from consideration. Recordings were made in an almost noise-free environment.

"Forensic cases usually have the worst noise situation," Klevans said. For example, a body recording device on an undercover agent in a train station will record not only the voices but also the background noise and even the agent's breathing. One of Klevans' goals is to render the system effective no matter how noisy the recording.

Another goal is to investigate the effect of disguised voices. "What we'd like to do is find those patterns in the voice that aren't changed by disguising or find mechanisms for detecting when someone is disguising their voice," Klevans said.

Discerning that a voice is disguised can prove innocence as well as guilt. "It may be that a squeaky voice sounds like an innocent bystander, and we can prove that it wasn't him by showing that it's some other guy disguising his voice," Klevans said.

According to Rosch, identifying voices quickly and accurately will save time and money, and will be an important tool in breaking large conspiracy cases involving drugs or guns. "It's work that is going to be very important for law enforcement," Rosch said.

The work is taking place in the Voice Input/Output Laboratory, directed by Dr. Robert Rodman, associate professor of computer science and primary investigator for the project.

/ News Index / News Archives Index /