Software that picks up voice patterns promises to provide a foolproof lie detector.
There are two types of detective. There are those convinced they can spot a lie from a tiny facial tic, a glance away or a sudden shift in the chair. Then there are those who never find themselves able to pick up on those signals from the body or, more importantly, link them to any admission of wrongdoing.
During his career as a police detective, Will Gaskell found he was never going to nail any villains by watching their body language. For him, it all came down to the questions - even though colleagues were sure they could spot a criminal.
"They had this view that you had to see the whites of their eyes," says Gaskell. "I was a career detective, but I never found too many visual cues that were all that useful. Things like crossing their arms or sitting forward, all those things that people think of when they believe someone is lying. But nobody really gives that much away in their body language."
It is not a surprise then that Gaskell was sceptical of the claims made by Nemesysco for its voice-analysis technology: software that is meant to pick up on subtle changes in speech that indicate whether someone is relaxed or showing stress from telling lies. He went to Israel, twice, to see the technology in action, and took along a psychologist to see if Nemesysco's claims were true. According to Gaskell, the psychologist told the computer a pack of lies: that he was a visiting alien and other nonsense. "It gave no indication of deception at all. It was clear that the kit could be deceived," claims Gaskell.
Yet insurance companies, police forces and government departments have bought into the technology, believing that it is an effective weapon against fraud. What is the software really doing? Nemesysco and its licensees (the Israeli company has arranged exclusive distribution agreements in a number of countries around the world) do not like to go into details about how the technology operates. But, like a number of other 'lie detectors', it concentrates on how people act when they are under stress.
The theory is that speech patterns change subtly in reaction to stress, and voice-analysis software looks for these patterns. To be sure, there is a lot of evidence that stress can alter the way people speak or move - but do those simple signals tell you anything useful? The subject may be stressed for any number of reasons.
University of Portsmouth researcher Aldert Vrij is deeply sceptical of the value of this kind of software. "It is as accurate as flipping a coin," he warns. "We know that it doesn't work and that it can't work."
"If users believe they are getting value out of using voice-risk analysis, it is probably because they have adjusted their way of working. In that view, the technology is, to a large extent, a placebo. It may be that they are digging deeper for people who fail the test," suggests Vrij, who also assures that subjects are aware that there is a machine in the background monitoring them. Therefore, they feel uncomfortable about proceeding with a claim and back out voluntarily before they are "found out".
Andy Nixon, the head of eSure's fraud department, points out that his company advertises the fact that they use it and, as a result, they get cases that go away.
"It is difficult to quantify why claimants drop out of the process. We tell people exactly what we are doing. We tell them we comply with the Data Protection Act of 1998, that calls are recorded and analysed using voice-risk analysis. In all the time we have used it, we have not had anyone say they didn't want to go through the interview," explains Nixon.
The company maintains that the use of the technology means speedier claims handling for most people. "It is a two-stage process. We have a primary call that is script-based. That initial interview takes about 20 minutes. At the end, it is classified as either high risk or low risk. A high-risk claim goes to a more experienced handler and that takes up to an hour, so they can look at it more deeply. Both the primary and the secondary interviews have the technology in the background."
Other companies in insurance have tried voice-risk analysis and decided against using it, or continue to use the technology on a limited basis. "There is a third-party company that does it for us. We don't promote that we use it a lot," claims a spokes-person for Admiral Insurance.
Others, such as Prudential, have hung back from using voice-risk analysis. "We don't use it at the moment and we have not trialled it. It is, to an extent, untested technology," says a spokesman.
Not all claims go through the system: today eSure uses the system primarily for motor-theft disputes. Before bringing the system in house (a number of the big UK insurers contracted out voice-risk analysis to see how it fared before deciding to adopt it internally or simply reject it) eSure compared its results with those of a technique called 'conversation management' favoured by Gaskell's company, VFM. It concentrates on the answers in the interview itself. Although there are subtle differences between the techniques, voice-risk analysis as used by eSure and conversation management are not all that different, says Nixon. It is important to have the right sort of interview technique, he stresses.
"The technology is not a silver bullet in any shape or form. We don't repudiate based on the technology. It is only when we have hard evidence that we do that. It is a case of knowing the questions and how to ask them. People will often lie by omission. So you focus in on those aspects of the conversation," says Nixon. The idea is that people find it hard to lie and, as they try to find alternative explanations that fit their earlier answers, their cognitive load goes up and it becomes harder to hide the stress. That is what the software should then pick up.
Vrij led a study funded by the Environmental Science Research Council into the use of visual and verbal cues to spot lies as an aid to crime detection. As Gaskell found in his police work, conventionally accepted visual and verbal cues turned out to be not all that useful. However, organising the interview to increase the cognitive load on the participants did prove fruitful, Vrij found. Studies of two voice-stress analysis systems, such as one performed by the University of Oklahoma last year, found that neither could detect lies about drug use among prisoners.
Speaking of voice-risk analysis, Gaskell says: "That kit will only operate properly if you have a managed conversation going on at the same time." In his view, if you have taken the trouble to train people in the art of managed conversation, there isn't much the software is going to tell you.
Nixon insists that the technology has paid off for eSure. "Voice-risk analysis compares favourably with all the other options that we looked at. And it is much quicker and slicker: so the genuine claimant gets dealt with in a one-stop process."
Gaskell counters: "Even though there might be some results. Do users know what it is that is producing those results?"
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.