Document Type

Theses, Ph.D

Rights

Available under a Creative Commons Attribution Non-Commercial Share Alike 4.0 International Licence

Publication Details

Successfully submitted for the award of Doctor of Philosophy (Ph.D.) to the Technological University Dublin, 2011

Abstract

The investigation of the emotional dimensions of speech is dependent on large sets of reliable data. Existing work has been carried out on the creation of emotional speech corpora and the acoustic analysis of emotional speech and this research seeks to build
upon this work while suggesting new methods and areas of potential. A review of the literature determined that a two dimensional emotional model of activation and evaluation was the ideal method for representing the emotional states expressed in
speech. Two case studies were carried out to investigate methods of obtaining natural
underlying emotional speech in a high quality audio environment, the results of which were used to design a final experimental procedure to elicit natural underlying emotional speech. The speech obtained in this experiment was used in the creation of
a speech corpus that was underpinned by a persistent backend database that incorporated a three-tiered annotation methodology. This methodology was used to comprehensively annotate the metadata, acoustic data and emotional data of the recorded speech. Structuring the three levels of annotation and the assets in a persistent backend database allowed interactive web-based tools to be developed; a
web-based listening tool was developed to obtain a large amount of ratings for the assets that were then written back to the database for analysis. Once a large amount of ratings had been obtained, statistical analysis was used to determine the dimensional
rating for each asset. Acoustic analysis of the underlying emotional speech was then carried out and determined that certain acoustic parameters were correlated with the activation dimension of the dimensional model. This substantiated some of the
findings in the literature review and further determined that spectral energy was strongly correlated with the activation dimension in relation to underlying emotional speech. The lack of a correlation for certain acoustic parameters in relation to the evaluation dimension was also determined, again substantiating some of the findings in the literature.
The work contained in this thesis makes a number of contributions to the field: the development of an experimental design to elicit natural underlying emotional speech in a high quality audio environment; the development and implementation of a
comprehensive three-tiered corpus annotation methodology; the development and
implementation of large scale web based listening tests to rate the emotional dimensions of emotional speech; the determination that certain acoustic parameters are correlated with the activation dimension of a dimensional emotional model in
relation to natural underlying emotional speech and the determination that certain acoustic parameters are not correlated with the evaluation dimension of a twodimensional emotional model in relation to natural underlying emotional speech.

DOI

https://doi.org/10.21427/D7GK59

Download


Share

COinS