Research Projects

The long term goal of the ICISLT program is to dramatically improve the performance of speech and language systems.  To achieve these goals, we make a deliberate effort to remove many of the constraints that limit most speech and language research.  Research projects are not limited to techniques designed to give immediate improvement; some may take several years before they have progressed enough to show any improvement at all.  Methodologies are not constrained by the amount of processing power or amount of memory required.  Examples of on-going or proposed projects include:

• Knowledge and Data Acquisition
  • Million Hour Audio Corpus
  • Universal Dictionary
  • Petabyte Language Knowledge Base

   

• System Architecture and Training Methods
  • System of Jointly Optimized Multiple Modules
  • Training on Untranscribed Data

   

• Acoustic Modeling
  • Verification-by-Synthesis
  • Speaker Variability and Dialect Manifold Modeling

   

• Language Modeling
  • Acoustically-Sensitive Language Modeling
  • Injecting Lexical Syntactic and Semantic Preferences in Parser-based Models

   

• Scaling and Data Size
  • Scaling and Distributed Computing
  • Utilizing Large Quantities of Data and Knowledge
  • Knowledge Acquisition from Large Number of Informants

   

• Rescoring and Optimization
  • Speaker Variability and Dialect Manifold Modeling
  • Minimum Error Rate as a Constrained Optimization Problem

Please contact us if you wish to participate in or sponsor any of these projects.

Copyright © 2005 James K. Baker