The aim of our research is to acquire and implement the acoustic-phonetic knowledge of the Dutch language which is necessary for the automatical generation of a phoneme lattice from the acoustical speech signal. This phoneme lattice, generated by the acoustical preprocessor, can be used as input for the lexical module of a speech recognizer. The preprocessor is implemented in Prolog, so as to facilitate the representation of the knowledge in a rule-based system.

In this paper we will present rules used for recognition of phoneme classes, together with the formalism used for implementing the rules. In addition, some recognition results will be presented. As recognition of classes is performed independently, i.e. one recognition step for each class, recognition results (segmentlists) are obtained for every class. Calculation of overall performance involves a decision mechanism, which we will also discuss, to combine the segment-lists of individual phoneme classes to a lattice of phoneme classes, using the measures of certainty associated with the elements in the segment-list, and different types of knowledge, e.g. phonological and statistical knowledge.