This paper describes an experimental approach to continuous speech recognition being carried out at IBM Research. It is based on statistical methods that allow automatic extraction of recognizer parameter values from speech data. Research results are presented for both artificial and natural recognition tasks.
