Speaker verification with discriminative vector quantization

Abstract

Vector Quantization (VQ) has been used widely in the areas of data compression and digital signal processing. Speaker identification and verification systems based on VQ have also been proposed and tested. The thesis describes a Discriminative Vector Quantizer for speaker verification that trains the VQ codebook using samples from both the targeted speaker and competing speakers. By taking into account the phonetic characteristics of the competing speakers, error reduction of over 80 reduction for verification with unseen speakers using Discriminative VQ.

Date of Award	1998
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology