PhD Oral Exam - Ali Sarafnia, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Speech perceptual quality is inherently multi-dimensional, encompassing intelligibility, naturalness, distortion, and clarity, each of which reflects a different listener priority and a different set of acoustic correlates. Speech naturalness is a quality of speech that makes it sound as spoken by a normal human-being. Naturalness of a speech signal is important since it requires less cognitive effort to understand it. Speech naturalness arises from a combination of spectral consistency, prosody, and temporal dynamics, in which listeners integrate smooth formant trajectories, appropriate pitch and amplitude modulation, and predictable temporal structure in order to judge a voice as human-like. Yet, the existing speech quality metrics, are not sensitive enough to this aspect of speech quality. These metrics mainly check as to how clear the speech is, or how good it sounds overall. None of them directly measures spectral consistency, the smooth, natural pattern of sound frequencies that makes a voice sound more like that of a human-being rather than non-natural. Naturalness of speech is also important in many applications such as in hearing aids, speech recognition and speech synthesis applications.

The objective of this thesis is to propose a metric that takes speech naturalness into account, and to investigate the application of such a metric in the design of speech systems, particularly in the presence of different types of noise.

For the first part of the objective, a new metric called the spectral entropy-based metric (SEM) is proposed. It is based on Shannon entropy that measures how ordered or random a signal is. It works by comparing the entropy of a speech signal before and after it passes through a speech system. SEM is validated experimentally using the fixed-beam and the steerable-beam first-order differential microphone arrays, using speech samples from the TIMIT database. For the second part of the objective, it is shown that SEM apart from it being used as a quality measure, it can help in the design of a very-closely-spaced first-order differential microphone array, even in the presence of diffuse noise or point-source noise when the existing method of designing by maximizing the array gain fails.