Pryathna Sankaranarayanan | May 9, 2017 | 5 minutes
One of the key critical processes when deploying a voice biometric system is setting the threshold that separates true speakers from impostors. The logic is simple: if the voice biometric score exceeds the threshold the speaker is assumed to be the true speaker and if below, they are an impostor.
Traditionally, setting a threshold has been a “hit-or-miss” affair. It is usually left to the system integrator to adjust a “single global threshold” parameter that is applied across the total database of enrolments based on the security/convenience trade-off required for the business application. The problem is, set the threshold too high and a large number of true speakers are rejected causing usability and business issues; set the threshold too low risks high numbers of false acceptances and impostor break-ins.
But the problem is more profound than that. Setting a “single global threshold” across the whole voice print database assumes that all voice prints exhibit the same security performance. Whilst it is true that voice prints created in controlled laboratory conditions exhibit relatively uniform performance; in “The Real World” noise, interference, distortion and speaker behaviour and variations in telecommunications channels often lead to voice prints with widely varying security performance. Consequently, a fixed global threshold does not compensate for these variations and leads to a sub-optimal solution for “Real World” commercial deployments.
Auraya’s ArmorVox overcomes this problem by, instead of using a single global threshold, it utilises thresholds set for each individual voice print enrolled in the system. Using speaker adaptive core voice biometrics technology developed by Auraya, instead of a “hit-or-miss” approach, ArmorVox measures the security performance of each voice print enrolled in the system and adjusts the threshold for each individual voice print based on the security performance of that voice print. This way ArmorVox compensates for variations in individual voice print performance and providing an optimal solution; irrespective of noise, interference, distortion, channel characteristics and variations in individual speaker behaviour.
As well as delivering an optimal solution, Auraya’s approach to individual voice print threshold settings also has an additional security benefit. The “single global threshold” represents a “Single Point of Vulnerability (SPV)”. Developed as a concept for assessing risk in mission critical infrastructure, SPV analysis assesses potentially catastrophic security problems by identifying single points of vulnerability in a system or solution. In voice biometrics the “single global threshold” represents one such SPV. Compromising this single global threshold comprises the security of ALL assets protected by the voice biometric system. However, in ArmorVox, given that each voice print has its own threshold; only those assets protected by that individual voice print are exposed if that threshold is compromised.
As a consequence, Auraya’s speaker adaptive core voice biometrics technology not only delivers an optimal solution, it is also the preferred technology for mission critical applications in telecommunications, banking, financial services, Government and national security.