Title page for ETD etd-08062003-112127 ( Browse

Type of Document

Dissertation

Author

Dognin, Pierre L.

Author's Email Address

dognin@siglab.ee.pitt.edu

URN

etd-08062003-112127

Title

A Bandpass Transform for Speaker Normalization

Degree

Doctor of Philosophy

Program

Electrical Engineering

School

School of Engineering

Advisory Committee

Advisor Name	Title
Dr. Amro A. El-Jaroudi	Committee Chair
Dr. C.C. Li	Committee Member
Dr. J. Robert Boston	Committee Member
Dr. Luis F. Chaparro	Committee Member
Dr. Mihai Anitescu	Committee Member

Keywords

Automatic Speech Recognition
Analytical Function
Bilinear Transformation
Feature Transformation
Frequency Warping
Front End Processing
Model Adaptation
Nelder-Mead Optimization
Non-Linear Transformation
Speaker Normalization
Vocal Tract Length Normalization

Date of Defense

2003-07-23

Availability

unrestricted

Abstract

One of the major challenges for Automatic Speech Recognition is to handle speech variability. Inter-speaker variability is partly due to differences in speakers' anatomy and especially in their Vocal Tract geometry. Dissimilarities in Vocal Tract Length (VTL) are a known source of speech variation. Vocal Tract Length Normalization is a popular Speaker Normalization technique that can be implemented as a transformation of a spectrum frequency axis. We introduce in this document a new spectral transformation for Speaker Normalization. We
use the Bilinear Transformation to introduce a new frequency warping resulting from a mapping of a prototype Band-Pass (BP) filter into a
general BP filter. This new transformation called the Bandpass Transformation (BPT) offers two degrees of freedom enabling complex
warpings of the frequency axis that are different from previous works with the Bilinear Transform. We then define a procedure to use BPT
for Speaker Normalization based on the Nelder-Mead algorithm for the estimation of the BPT parameters. We present a detailed study of the
performance of our new approach on two test sets with gender dependent and independent systems. Our results demonstrate clear
improvements compared to standard methods used in VTL Normalization. A score compensation procedure is presented and results in further
improvements of our results by refining our BPT parameter estimation.

Files

Filename		Size	Approximate Download Time (Hours:Minutes:Seconds)
Filename		Size	28.8 Modem	56K Modem	ISDN (64 Kb)	ISDN (128 Kb)	Higher-speed Access
	Dissertation_Pierre_L_Dognin_23July2003.pdf	778.45 Kb	00:03:36	00:01:51	00:01:37	00:00:48	00:00:04