CAREER: A Graphical-Model Based Software Infrastructure for Speech Recognition Research and Education

University Of Washington, Seattle WA

Investigators

Abstract

This is the first year funding of a 5-year continuing award. Current core speech recognition technology is inadequate in several important situations, where high error rates often render the technology unusable. This project will advance the state of the art through use of graphical models, a relatively new statistical technique comprising formal methods for reasoning about random processes, and sets of algorithms for making probabilistic data-driven decisions in computationally efficient ways. All existing statistical techniques used in automatic speech recognition (and many more) appear to be representable using graphical models; no other known abstraction has such a property. The PI will design and build new freely available open-source software for speech recognition research and education. This infrastructure will provide the tools needed to successfully investigate different graphical-model based speech recognition algorithms, and will include tools for both computing with a specified model and for searching for new ones. The infrastructure will be used for curriculum development at the University of Washington (both at the graduate and undergraduate level), and will be released, along with course materials, on the Internet for free use by anyone.