Continuous improvement in natural-language speech recognition (NLSR) technology and its transfer onto mobile devices are helping cement voice biometrics as a viable alternative mechanism for user authentication, a biometrics expert believes.
Building on statistical voice analysis techniques it acquired with its 2011 purchase of Italian voice-recognition specialist Loquendo, Nuance Communications has been expanding the scope of the technology away from conventional desktop voice recognition, and this year has been promoting new platforms such as its Nuance Forensics voice-matching technology and its Dragon Mobile Assistant app for Android.
“We've seen a great movement for mobiles as becoming the default device for interaction with the organisation,” the company's senior technical lead Sasha Agafonoff told CSO Australia, noting the increasing use of the technology to verify the identity of people not only when they use apps but also when they dial into contact centres.
“Anecdotal reports from large [banking] deployments overseas suggest that customers love the system, and find it much more secure and easier to access,” he continued.
“They feel their institution is protecting them by putting in these extra measures, and they're not having to expose themselves to solutions like PINs and passwords that are easy to forget and easy for others to guess.”
Some solutions have struggled “in getting the customer experience right", he conceded, noting that the registration and enrolment process for customers needs to be easy if the biometric solution is to work properly.
Yet with cloud-based statistical analysis allowing for the collection of increasingly large quantities of user voice information, accuracy for the solutions continues to increase – to the point where users are no longer wrestling with voice-recognition platforms as they used to.
“We now have such a huge volume of transcription happening now, that it means the amount of data available for us to make really significant improvements in the technology will have us seeing really good improvements over time,” Agafonoff said.
The use of multiple algorithms and probabilistic linear discrimination analysis techniques – as well as recent innovations such as iVectors, which adjust recognition for the speaker's environment – made it possible “to really consider the person's characteristics, as distinct from another person", he explained.
This ability – augmented by Loquendo's own technology and its incorporation into the new Nuance Forensics platform – had improved the technology enough that it was now accepted in court as evidence in proving that a recording of a particular voice belonged to a particular person.
“We're able to conduct a statistical analysis of a person's voice and to pull out those aspects of the voice that identify that person as distinct from a representative group of peers,” he explained. “There is a science behind producing these sorts of reports.
“Working with multiple biometric algorithms, we are able to produce much more accurate results – and that's delivering the kind of benefit that we're able to use to calculate likelihood ratios and chance residual error.
“This will help provision of speech technologies and speaker recognition technologies to improve customer experience – and we'll see this combined with federated authentication services that combine a range of factors to provide voice services to a range of institutions.”