Monday, November 22, 2010

Audio Forensics: Voice Risk Analysis Fails Practical Trial

Voice Risk Analysis (VRA) has always been viewed skeptically by many in the speech field. This technique takes Voice Stress Analysis - i.e. analyzing speech for microtremors and possibly other indicators of nervousness, anxiety, and general stress - and applies it to a real world problem. As far as the science of strict voice stress analysis goes, it is believable to me based on my own observations. I have seen that many people can exhibit audible stress in their voices when under emotional or physical stress.  As an experiment, try listening to someone on the phone when they are climbing a set of stairs rapidly and you should notice a difference!

Voice Risk Analysis takes this concept and extends it to the automated detection of deception (lying) over telephone calls.  Is this completely unbelievable on the face of it? Not completely - you may have occasionally encountered people who sound nervous when trying to lie.  However, pricey products have been sold based on extrapolating this a step further to claim that simply by analyzing telephone speech to a finer level of detail (e.g. looking for microtremors) one can detect stress and even lying in the general population.

The Guardian (a center-left news organization in the UK) has a news article on the results of practical trials conducted by the British government using a commercial VRA system to attempt to catch people lying in telephone interviews regarding government-funded benefits. After analyzing the results from 45,000 calls, the bottom line was that the system did not provide enough value to justify its cost.  Given the history of VRA and my own understanding of speech science, I am not surprised at all by the findings.  Hopefully, as a matter of public policy, this will put an end to government purchases of VRA technology until such a time as it is proven by peer-reviewed studies to work for the application(s) it is intended for.

No comments: