This thesis investigates a statistical approach to tracking formant trajectories in continuous speech. In this approach a probability measure is applied to a set of features extracted from each analysis frame of the speech signal, and a conditional mean estimate is used to determine formant frequency values. The features used can be vector quantization symbols, spectrum levels, or other sets of features related to formant frequencies. Continuity constraints can be applied via either simple smoothing algorithms or hidden Markov models. An example of this technique using a multivariate probability measure on LPC spectral values is examined in detail. A second example using vector quantization is also examined for comparison. The performance of these trackers under a variety of conditions is discussed.

Library of Congress Subject Headings

Speech processing systems--Design; Formants (Speech)--Statistics--Data processing; Automatic speech recognition

Publication Date


Document Type


Department, Program, or Center

Computer Science (GCCIS)


Hillenbrand, James

Advisor/Committee Member

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 G395 1989


RIT – Main Campus