Probabilistic random projections and speaker verification

Abstract

Biometrics is susceptible to non-revocable and privacy invasion problems. Multiple Random Projections (MRP) was introduced as one of the cancellable biometrics approaches in face recognition to tackle these issues. However, this technique is applicable only to ID fixed length biometric feature vector but failed in varying size feature, such as speech biometrics. Besides, simple matching metric that used in MRP unable to offer a satisfactory verification performance. In this paper, we propose a variant of MRP, coined as Probabilistic Random Projections (PRP) in text-independent speaker verification. The PRP represents speech feature in 2D matrix format and speaker modeling is implemented through Gaussian Mixture Model. The formulation is experimented under two scenarios (legitimate and stolen token) using YOHO speech database. Besides that, desired properties such as one-way transformation and diversity are also examined. © Springer-Verlag Berlin Heidelberg 2007.