High performance voice authentication system

• Main Author: Panda Ashish
• Other Authors: Srikanthan, Thambipillai
• Format: Theses and Dissertations
• Published: 2008
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
• Online Access: http://hdl.handle.net/10356/2420
• Institution: Nanyang Technological University

Extensive simulations were performed on two popular speech databases, namely KING and TIMIT, to evaluate the proposed methods. A new background model called Global Background Model (GBM) has been presented to replace the memory intensive Universal Background Model (UBM). Based on a novel set theoretic framework for UBM, it has been analytically shown that the performance of the GBM is comparable to that of the UBM. In the quest for efficient algorithms for model training, applicability of vector quantization algorithms for training a GMM has been studied. Subsequently a Bayes Adaptation (BA) based training scheme has been proposed to replace the iterative Expectation Maximization (EM) algorithm for rapid speaker model estimation. Experiments conducted on the speech databases reveal that BA based training scheme results in comparable, at times even better, accuracy as compared to the EM algorithm scheme, while significantly reducing the training time.