We have investigated the benefits of an interaction between lexical knowledge and acoustic analysis in an isolated word recognizer. By using lexical knowledge, the need for detailed acoustic analysis can be reduced and higher recognition rates can be obtained. In our proposal the lexical knowledge we use is determined by hierarchies of phonetic classes. These hierarchies have the form of binary decision trees and are used to guide the recognition process. A description of the proposed recognition process is given.
By means of a simulation of this recognition process we are able to predict word recognition and confusion rates for the recognition system. Results of this simulation for different Dutch lexica are given.