Pryathna Sankaranarayanan | December 8, 2017 | 4 minutes
At Auraya, we design solutions that help businesses manage risks. Lately, this includes voice imitation, synthetic speech and playback attacks.
Start-ups such as Lyrebird have been in the news for a technology that creates digital voice from existing audio files. The technology is impressive and has numerous applications. However, it can be a potential threat to voice biometrics deployments.
ArmorVox advanced security delivers a robust and reliable technological solution to manage fraud by addressing this risk.
A major difference between natural voice and pre-recorded voice is that natural voice varies slightly – even when the individual is repeating the exact phrase or digits a few times. A pre-recorded voice sounds the exact same each time.
Tone, pitch, speed, and memory and attention of an individual while speaking makes natural voice vary each time. Such changes are expected and are actively learned by the underlying machine learning algorithms of ArmorVox.
On the other hand, if the voice sounds ‘too perfect’, the technology flags this as a potential fraud and increases the security setting. Increased security setting may mean a random digit string that appears on the speaker’s device that needs to be said by the enrolled speaker immediately, thus reducing the capability of a fraudster to use pre-recorded or synthetically created responses.
The Random Challenge Response is a unique digit sequence that appears on the speaker’s device for enhanced security. This ensures liveness testing – that along with an individual’s voice, they can speak the correct digits displayed on the device to be successfully verified. This is an additional security setting that can be raised when the technology detects a possible fraudulent attempt such as a playback attack.
To ensure sounds such as tones, musical instruments, noise and synthetic voice are not used to enrol a voice print, a sound detection model tracks every voice enrolment and only allows human generated voice to be enrolled.
The sound detection model is designed with underlying characteristics of natural voice which is distinctly different to artificial voice. It is sensitive to other sounds that are artificially created and alerts organisations of fraud in real-time.
With new and emerging security threats, the team at Auraya can design voice models that capture distinct sound characteristics of the threat – and integrate the model to ArmorVox to differentiate between true speakers and fraudulent attempts. With our security design, we ensure that businesses stay a step ahead in managing risk and have sustainable business outcomes by implementing voice biometrics.