This thesis investigates a stochastic modeling approach to word hypothesis of phonetic strings for a speaker independent, large vocabulary, continuous speech recognition system. The stochastic modeling technique used is Hidden Markov Modeling. Hidden Markov Models (HMM) are probabilistic modeling tools most often used to analyze complex systems. This thesis is part of a speaker independent, large vocabulary, continuous speech understanding system under development at the Rochester Institute of Technology Research Corporation. The system is primarily data-driven and is void of complex control structures such as the blackboard approach used in many expert systems. The software modules used to implement the HMM were created in COMMON LISP on a Texas Instruments Explorer II workstation. The HMM was initially tested on a digit lexicon and then scaled up to a U.S. Air Force cockpit lexicon. A sensitivity analysis was conducted using varying error rates. The results are discussed and a comparison with Dynamic Time Warping results is made.

Library of Congress Subject Headings

Automatic speech recognition; Phonetics, Acoustic--Analysis--Data processing; Natural language processing (Computer science); Hidden Markov models

Publication Date


Document Type


Department, Program, or Center

Computer Science (GCCIS)


Biles, John

Advisor/Committee Member

Anderson, Peter


Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TK7882.S65 E536 1990


RIT – Main Campus