Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Hybrid Threshold Approach in Text-Independent Speaker Verification

Fangxin Chen, Bruce Millar, Michael Wagner

TRUST Project, Research School of Information Sciences and Engineering Australian National University

In this paper we suggest a hybrid threshold approach for text-independent speaker verification. A maximum distortion (or minimum likelihood) threshold is set for the claimed speaker to perform pre-filtering of the highly dissimilar impostors. Cohort normalisation then is applied to further separate those impostors who are acoustically similar to the claimed speaker. A VQ-distortion based text-independent speaker verification system using this approach achieves better results than the conventional absolute threshold or cohort normalisation methods.

Full Paper

Bibliographic reference.  Chen, Fangxin / Millar, Bruce / Wagner, Michael (1994): "Hybrid threshold approach in text-independent speaker verification", In ICSLP-1994, 1855-1858.