Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Speaker Recognition with Discriminative Speaker VQ Models

Kai Tat Ng (1), Jian Su (2), Bingzheng Xu (1)

(1) Department of Electronic Engineering, City University of Hong Kong, Hong Kong
(2) Institute of Radio and Automation, South China University of Technology, China

In this paper, we propose a discriminative VQ model for speaker recognition. The VQ training algorithm is developed within the framework of gradient descent learning. Unlike the conventional VQ scheme, which considers only the minimum distance to a speaker codebook, the new algorithm takes account of the distances to all competing classes and all codewords in a speaker codebook, and aims directly at minimizing the recognition error. Two sets of speaker recognition experiments based on the conventional learning scheme and the DVQ are conducted respectively on a database of 200 French speakers. We obtained 1.75% performance improvement in speaker recognition and 0.45% in verification over the conventional VQ algorithm.

Full Paper

Bibliographic reference.  Ng, Kai Tat / Su, Jian / Xu, Bingzheng (1995): "Speaker recognition with discriminative speaker VQ models", In EUROSPEECH-1995, 325-328.