Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

A Priori Threshold Selection for Fixed Vocabulary Speaker Verification Systems

Arun C. Surendran, Chin-Hui Lee

Multimedia Communications Research Lab, Bell Labs, Lucent Technologies, Murray Hill, NJ, USA

A priori threshold selection is an important problem in practical speaker verification (SV) systems. Most earlier empirical methods for estimating prior thresholds assume availability of data from impostors i.e. speakers saying the same phrase as the desired speaker. While this is true for most databases available for research, it is not so in practice. In this paper,we present a novel prior threshold selection mechanism that uses the distributions of the "units" of the pass phrase rather than the entire phrase. Hence the new approach does not need imposter data and provides flexibility and adaptability to the system. We demonstrate the effectiveness of our new idea in a connected-digit based SV system.


Full Paper

Bibliographic reference.  Surendran, Arun C. / Lee, Chin-Hui (2000): "A priori threshold selection for fixed vocabulary speaker verification systems", In ICSLP-2000, vol.2, 246-249.