Probabilistic Random Projections and Speaker Verification
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 1D 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.
KeywordsSpeaker verification Cancellable biometrics Probabilistic Random Projections
- 3.Soutar, C., Roberge, D., Stoianov, A.R., Gilroy, Vijaya Kumar, B.V.K.: Biometrics Encryption. In: Nichols, R.K. (ed.) ICSA Guide to Cryptography, pp. 649–675. McGraw-Hill, New York (1999)Google Scholar
- 5.Savvides, M., Vijava Kumar, B.V.K., Khosla, P.K.: Cancelable Biometrics Filters for Face Recognition. Int. Conf. of Pattern Recognition 3, 922–925 (2004)Google Scholar
- 8.Chong, L.Y., Teoh, B.J.A., Khor, S.E.: 2D CLAFIC Subspace Technique in Probabilistic Speaker Verification. In: Fifth IEEE Workshop on Automatic Identification Advanced Technologies, IEEE Computer Society Press, Los Alamitos (2007)Google Scholar
- 9.Teoh, B.J.A., Chong, T.Y.: Cancellable Biometrics Realization with Multispace Random Projections. Accepted to publish in IEEE Transactions SMC Part B (2007)Google Scholar
- 12.Higgins, A.: YOHO Speaker Verification. In: Speech Research Symposium (1990)Google Scholar