Text-independent speaker verification based on vector quantization

Pao Yue-kong Library Electronic Theses Database

Text-independent speaker verification based on vector quantization

 

Author: Wong, Chuen-kau
Title: Text-independent speaker verification based on vector quantization
Degree: M.Sc.
Year: 1998
Subject: Speech processing systems
Automatic speech recognition
Hong Kong Polytechnic University -- Dissertations
Department: Multi-disciplinary Studies
Dept. of Electronic Engineering
Pages: iv, 52 leaves : ill. ; 30 cm
Language: English
InnoPac Record: http://library.polyu.edu.hk/record=b1436950
URI: http://theses.lib.polyu.edu.hk/handle/200/322
Abstract: This thesis implements text-independent speaker verification experiments for clean and telephone speech based on vector quantization. The well-known speech databases, TIMIT and NTIMIT, are used. 12th-order LPC-derived cepstrum coefficients are extracted from 50% overlapping frames of speech signal and are used as input feature vectors. Each speaker in the speaker set is modeled by a personalized VQ codebook and is associated with a decision threshold. Each speaker's threshold is estimated by the distributions of the intraspeaker and interspeaker distances, which are assumed Gaussian. All thresholds are then approximated by a linear regression line. Cepstral distance measure is naturally chosen for comparing the similarity of two spectral vectors. Full search Generalized Lloyd algorithm with Binary Splitting method is used to train the codebooks. The codebook sizes are varied from 16 to 128. The effects of different codebook sizes on system performance are demonstrated. The first set of experimental results show that codebook size of 64 performs the best for both clean and telephone speech. Average verification error rates of 1.01% and 22.16% have been achieved respectively for clean and telephone speech. The second experiment evaluates the effects of using the index-weighted cepstral distance measure on the verification performance. As expected, better results are obtained for high quality speech. Furthermore, codebook size of 128 produces a fairly low average error rate of 0.61% but no improvement is observed for noisy channel. The final experiment demonstrates the effect of using combined codebooks on the system performance. In addition to the instantaneous VQ codebooks, a transitional VQ codebook, which contains delta-cepstral vectors as its codebook entries, is added to model each speaker. The relative contribution of instantaneous and transitional spectral information is investigated. The optimal combination factors for codebook sizes of 64 and 128 are found empirically. An average error rate of slightly less than 1% for codebook size of 64 has been obtained for clean speech but poor results for telephone speech still exist. All these lead to a conclusion that pre-processing of telephone speech signal for removing the noise and channel variations is necessary.

Files in this item

Files Size Format
b14369503.pdf 2.057Mb PDF
Copyright Undertaking
As a bona fide Library user, I declare that:
  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

     

Quick Search

Browse

More Information