Automatic Speech Recognition Using Morpheme Structure Rules for Word Hypothesis and Dictionary Generation

The object of this thesis is to investigate the use of Morpheme Structure Rules as part of a Knowledge Based Speech Recognition System. The prescription and use of such rules is presented within the framework of a metatheory developed as a tool to assess the design and development of Speech Recognition Systems as a simulation of the speech perception mechanism. The PROLOG programming language was used throughout as an implementation vehicle for the simulations. The input to the system is in the form of broad class phonological categories with a number of phonological unit labels attached to each class. The acoustic labelling front end of the system was simulated using data obtained from hand labelled spectrograms. Both isolated and continuous, monosyllabic, word input were simulated. The morpheme as a unit of recognition is compared with the syllable and the word. A learning dictionary formed part of the system, which enabled the recognizer to tune to a particular vocabulary, with minimal user interaction. Results show that Morpheme Structure Rules provide a valuable source of knowledge when trying to determine sequences of phonological units for further processing and allow the capability of a learning dictionary.