CRA Digital Government Fellows Program

THIS IS AN ARCHIVED VERSION OF CRA'S WEBSITE. THIS ARCHIVE IS AVAILABLE TO PROVIDE HISTORICAL CONTENT.

PLEASE VISIT HTTP://WWW.CRA.ORG FOR THE LATEST INFORMATION

About CRA

Membership

CRA for Students

CRA for Faculty

CRA-Women

Computing Community Consortium (CCC)

Awards

Projects

Events

Jobs

Government Affairs

Computing Research Policy Blog

Publications

Data & Resources

CRA Bulletin

What's New

Contact

Home

CRA Digital Government Fellows Program

Back to Digital Gov't Fellows home page

"Graphical Models for Speech Recognition"
Jeff Bilmes, University of Washington
June 11th, 2001
Federal Communications Commission, Washington DC

Jeff Bilmes, an assistant professor at the Department of Electrical Engineering at the University of Washington, Seattle, gave a lecture on Graphical Models for Speech Recognition.

Professor Bilmes is also an adjunct assistant professor in linguistics. He co-founded the Signal, Speech, and Language Interpretation Laboratory at the university. He received a Masters degree from MIT, and a Ph.D. in Computer Science at the University of California in Berkeley in 1999. His primary research interests lie in statistical modeling (particularly graphical model approaches) and signal processing for speech and pattern recognition, and language and audio processing.

"Graphical Models for Speech Recognition"

A signal processing technique, graphical models (GMs) are flexible statistical abstractions that offer a promising path on which to find new approaches to automatic speech recognition (ASR).

This talk provided a brief overview of GMs, covering the four main components (semantics, structure, implementation, and parameters) needed to associate a given GM with a probabilistic model. Three types of graphical models were discussed. The first type are used to represent the process of classifier combination yielding novel combination rules that have improved accuracy for voice recognition in a phonetic classification task over previous combination schemes. A second type of GMs, known as hidden-articulator Markov models are used to represent hidden information about articulatory gestures within the human vocal-tract. Finally, buried Markov models (BMMs), are used where dependencies in a HMM have been augmented in discriminative and data-derived ways. This is a useful property for a GM when used for the classification task (e.g., ASR). Relative to other generative models, such "discriminative-generative" models have the promise to improve parsimony (i.e., have smaller memory and compute demands), yet improve recognition accuracy.