Date of Award

Spring 2003

Document Type

Thesis - Restricted

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering

First Advisor

Johnson, Michael T.

Second Advisor

Heinen, James A.

Third Advisor

Povinelli, Richard J.

Abstract

One of the key issues in speech recognition is to find a good acoustic model which accounts for important variabilities, like co-articulation, that affect speech. From all the models developed to reach this objective, the triphone model is the most successful one. Triphones are typically estimated. through training. Due to that fact, triphones are subject to limitations such as the need for a large set of balanced training data, and the impossibility to have examples needed for all possible models. Our goal in this research is to understand how co-articulation works and to formulate a model for co-articulation, which will be used to create triphones directly from monophones. To achieve this interesting but challenging goal, we model triphones as an interpolation between monophornes, and derive the rules that govern this interpolation. We call this method "Triphone creation through rule-based trajectory interpolation".

Share

COinS

Restricted Access Item

Having trouble?