A simulation study of the usefulness of broad phonetic classification in automatic speech recognition

The authors present a method to develop a hierarchy of phonetic classes which describes the set of subclasses which allows the identification of all words in a lexicon, with a minimum of detailed acoustic information. Resulting hierarchies for Dutch lexica are given. Such hierarchies show at what points it is most efficient to improve an automatic speech recognition system from a lexical point of view. When the acoustic performance of an automatic speech recognition system is known in terms of a phoneme confusion matrix, the analyses presented permit prediction of the optimal performance of the system for isolated words, depending on the lexicon to which the recognizer is applied