This paper proposes two methods for creating a pooled model for all registered speakers to reduce the enormous amount of calculation needed by the similarity normalization method for speaker verification based on a posteriori probability. The proposed methods perform the same as or better than the original method and the amount of calculation is reduced significantly. Speaker verification is tested by using separate populations of customers and impostors in order to evaluate performance under practical conditions. The speaker (and text) verification error rates are roughly 1.6 times larger than if the same population is used for both customers and impostors. Using IS customers and a separate group of 15 impostors, one proposed method achieves a speaker verification error rate of 1.6% for text-independent verification and a speaker and text verification error rate of 1.1%, which is about half that with the original method in text-prompted verification.
Cite as: Matsui, T., Furui, S. (1994) Similarity normalization method for speaker verification based on a posteriori probability. Proc. ESCA Workshop on Automatic Speaker Recognition, Identification and Verification, 59-62
@inproceedings{matsui94_asriv,
author={Tomoko Matsui and Sadaoki Furui},
title={{Similarity normalization method for speaker verification based on a posteriori probability}},
year=1994,
booktitle={Proc. ESCA Workshop on Automatic Speaker Recognition, Identification and Verification},
pages={59--62}
}