Can voice analytics detect a lie?
31 May 2019
“Fake News” is not just on the Internet, it is all around us. Whilst some people use white lies to make their life easier, some use them to manipulate a situation or individual and in extreme cases use them to commit crime. In this article, I will be exploring how voice analytics technology can be used to identify fraud.
The world of poker is an interesting place to observe the practice of lying. Players are often required to interpret human behaviour in order to detect if a player is lying, or in poker terminology “making a bluff”. Due to the diverse and varied nature of human behaviour, it is almost impossible to identify particular traits in every individual. However, there are some behaviours which may suggest a player is lying, such as a change in tone in their voice which could suggest the opponent is stressed. By correctly weighing and understanding behaviour, players are able to improve their success rate of identifying a bluff. It is the ability to find these edges and implement them consistently over the long term which will make the difference between the winners and losers.
Whilst poker relies on both physical and verbal traits to detect a lie, voice analytics technologies are only able to measure characteristics such as tone, tempo and arousal which are then mapped to emotional states. Similar to a game of poker, the theory is that when people are lying, they will exhibit higher signs of stress and anxiety.
There are a variety of business areas with use cases; from the monitoring of traders, to stopping identity theft, to detecting false insurance claims. The Association of British Insurers estimated that £1.28bn worth of fraud was detected in 2017. We can only imagine the difference that could be made by improving detection by a few percent.
If we return to the insurance use case, it is with the above method that frontrunners are currently utilising to assess the validity of claims. They are analysing recordings looking for baselines and trying to identify parts of the calls with vocal links to lying. Those which have a higher correlation to their model will then be flagged. There is much contention over the accuracy of these systems and I would recommend being sceptical of claims estimating 90% accuracy. In reality you are more likely looking at an accuracy of around 65% on fraudulent claims (the same as a trained human). However, the benefit is shown through the volume of material you can analyse and flag. As a result of this, more thorough investigations can be launched which can lead to identifying and stopping false claims.
For both insurance claims, and the wider area of fraud detection, voice analytics is not a fully polished solution quite yet. Practically there needs to be an improvement of accuracy and a reduction of false positives.This can come from an improvement in the models, but also by combining it in a system which includes analysis from many technologies that gathers information from claim histories, keyword analysis or in the case of insider financial crime, transaction monitoring.This is a rapidly developing space which is going to develop over the coming years and we think that voice analytics could be part of the wider solution.
There are also ethical implications which the industry will need to tackle. We should always remember that correlation does not necessarily mean causation, and while vocal indicators can suggest dishonesty, they can also correlate to other stresses and anxieties. There needs to be a rigorous structure in place, as it could be deemed immoral to flag someone for essentially acting in a human way. Further, as the technology improves there will need to be ethical controls to ensure that its use remains within its original remit and not in creating a Big Brother society.
The future of voice technologies at PwC...
Analysing voice tone for truthfulness is not the only way to detect fraud with voice technologies. There have been a lot of advances in technologies to help with fraud by impersonation around digital voice prints. Not only are some using this as a layer in the security process of verifying a caller, but it is also possible to create databases of voice IDs which have a high number of claims, or to identify if one voice is making multiple claims under different names.
Here at PwC, we are working with voice analytic vendors and combining their technical expertise with our industry knowledge: initial outcomes are positive and we are excited about the potential applications. If you’re interested in what emerging voice technology can do for you, then please get in touch. In my next blog I am going to talk about further use cases of voice technology, including how to evaluate your customer service and how the voice could be used to overcome the Acquiescence Bias in market research, which is where people have the tendency to agree with questions posed.