Statistical Recursive Estimation Algorithms for Speaker Adaption

The major thrust of this thesis is the development of statistical recursive estimation algorithms for speaker adaptation. In this thesis, a unified framework for recursive maximum likelihood estimation and recursive Bayesian learning is first developed and applied to direct hidden Markov model parameter estimation. Then an online Bayesian learning technique is proposed for recursive maximum a posteriori estimation of tree-structured linear regression and affine transformation parameters. This technique has the advantages of accommodating flexible forms of transformation functions as well as prior probability density functions. To balance between model complexity and best goodness of fit to adaptation data, an efficient dynamic programming based pruning algorithm is developed for selecting models using a Bayesian variant of the minimum description length (MDL) principle. Speaker adaptation experiments using a 26-letter English alphabet vocabulary were conducted, and the results confirmed effectiveness of the on-line learning framework.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Wang, Shaojun

Contributors dc:contributor

• Yunxin Zhao

• Engineering, Electronics and Electrical

Language dc:language

eng

Identifier

(MiAaPQ)AAI3017245

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/80724