Privacy-preserving iVector-based speaker verification

This work introduces an efficient algorithm to develop a privacy-preserving (PP) voice verification based on iVector and linear discriminant analysis techniques. This research considers a scenario in which users enrol their voice biometric to access different services (i.e., banking). Once enrolment is completed, users can verify themselves using their voice-print instead of alphanumeric passwords. Since a voice-print is unique for everyone, storing it with a third-party server raises several privacy concerns. To address this challenge, this work proposes a novel technique based on randomisation to carry out voice authentication, which allows the user to enrol and verify their voice in the randomised domain. To achieve this, the iVector based voice verification technique has been redesigned to work on the randomised domain. The proposed algorithm is validated using a well known speech dataset. The proposed algorithm neither compromises the authentication accuracy nor adds additional complexity due to the randomisation operations.