Date of Degree

9-2021

Document Type

Dissertation

Degree Name

Ph.D.

Program

Speech-Language-Hearing Sciences

Advisor

Douglas H. Whalen

Committee Members

Hosung Nam

Wei-rong Chen

Mark K. Tiede

Christina Hagedorn

Subject Categories

Phonetics and Phonology

Keywords

vowel variability, speaking rate, the uncontrolled manifold, UCM, normalizing flow, invertible neural networks

Abstract

Variability is intrinsic to human speech production. One approach to understand variability in speech is to decompose it into task-irrelevant (“good”) and task-relevant (“bad”) parts with respect to speech tasks. Based on the uncontrolled manifold (UCM) approach, this dissertation investigates how vowel token-to-token variability in articulation and acoustics can be decomposed into “good” and “bad” parts and how speaking rate changes the pattern of these two from the Haskins IEEE rate comparison database. Furthermore, it is examined whether the “good” part of variability, or flexibility, can be modeled directly from speech data using the flow-based invertible neural networks framework. The application of the UCM analysis and FlowINN modeling method is discussed, particularly focusing on how the “good” part of variability in speech can be useful rather than being disregarded as noise.

Share

COinS