Adapting for Age: The Limitations of Modern Voice Biometric Technology

You’d think that the signs of ageing are obvious. Grey hairs, wrinkles and fading eyesight. But, as a study by Pindrop points out, it’s also your voice that can change.
With the growing use of voice recognition as a way to identify and authenticate callers, there has been a real need to make sure that businesses using this technology are doing so accurately. False readings not only hinder the customer experience, they can also open the door for fraudsters. Age, it turns out, can be a huge factor here.
A changing voice
Pindrop’s recent two-year study into voice ageing analysed 122 people, including native English, Dutch, German, Spanish and Italian speakers. It found that the expected error rate (EER) of positively identifying a speaker increased as time passed and the survey sample aged. In fact, in the study, the EER almost doubled over the course of the two-year period. It demonstrates that organisations who depend on voice biometric technology may find it very difficult to authenticate a person on voice alone, especially if they are an infrequent caller, as time passes.
The research also revealed that other factors, besides age, can lead to a change in voice. Dr. Elie Khoury, the principal researcher in the study, says that a person’s emotional state, stress levels, health and vocal effort can affect the pitch and speed of their voice, and therefore the accuracy of identification. What’s more, someone calling from a mobile phone in a rural part of Germany will sound different from someone calling from a landline in Berlin – a subtle difference that may not always be picked up by standard voice biometric technology.
Finally, the research suggests that people’s voices age uniquely at different rates. Which means that there is no one accepted factor that can be applied to take ageing into account. This finding highlights one of the key flaws of voice biometrics: it’s trouble with adapting to voice variation. Because unlike irises or fingerprints, which stay the same over the course of a person’s life, a changing voice can directly affect the accuracy of acceptances or rejections.
“Voice biometrics aren’t accurate enough on their own,” Khoury says. “You have to add other factors like spoofing detection and Phoneprinting™.”
Hear this out
Compensating for the changing properties of an individual voice will be the main challenge researchers have when they are looking to improve voice recognition in the future. But it should not be a deal breaker for companies considering implementing voice technology.
According to Dr. Khoury, updating a model frequently enough can account for voice ageing – a theory he tested on 400 recordings of Barack Obama’s public speeches, taken from the inauguration in 2009 up until January 2017. By recalibrating the biometric model, he significantly reduced the effects that time had on the score.
Yet, as Khoury points out, the idea is not without risk, unless properly protected. “You can update the model with each new recording,” he says. “But that’s risky if someone is able to attack the system and compromise it.”
His warning is simple. Organisations shouldn’t rely on voice biometrics for authentication alone. Rather, they should embrace a multi-layered approach to security, considering safeguards such as Phoneprinting™. Because while voices may change, the need to balance customer service and fraud prevention will always remain.
Learn to Balance Customer Service and Fraud Prevention