Richard Smith

Word Hypothesization for Large-vocabulary Speech Understanding Systems Degree Type: Ph.D. in Computer Science
Advisor(s): Lee Erman
Graduated: May 1978

Abstract:

This thesis describes research directed toward the development of a general English speech understanding system. In particular, the thesis presents the design and performance of a bottom-up word hypothesizer Noah capable of handling very large vocabularies. The design of Noah is based on a hierarchy-tree structure. Speech is represented at four levels of a hierarchy. A tree maps the representation of speech at one level to the representation of speech at the next higher level by a tree.

The author concludes that bottom-up word hypothesization is not greatly effected by the size of the vocabulary. He was pleasantly surprised that the effect of vacabulary size on performance and on computation costs would be approximately according to the logarithmic of the vocabulary size. This result suggests that, with improvements in the word hypothesizer and the segmenter-labeler, speech understanding systems for general English can obtain a great amount of constraint from the acoustics alone.

Thesis Committee:
Lee Erman (Chair)

Joseph Traub, Head, Computer Science Department