The Use of Formal Grammars in Automatic Speech Recognition

Abstract

An automatic isolated-word recognition (IWR) system normally consists of a feature extractor (FH) followed by a recognition processor. Some form of 'training' is usually required in order to combat problems of variations in speech. This thesis presents the application of formal grammars to model a FE in an IWR system. The method is to construct, in the training mode, one grammar for each word in the vocabulary, directly from a set of sample strings of 'features' represented by symbols. In the recognition mode, an incoming string is analysed to determine which grammar, if any, could have generated it. Inference algorithms for both finite-state grammars (FSG's) and context-free grammars (CFG's) considered here are based on the eriterion of maximizing the similarity between various strings of the same word. The classification of a string involves the use of the ‘weighted matching network' technique in the FSG approach and the computation of the minimisation matrix M for the CFG approach. Both the FSG and CFG models offer comparable recognition performances whilst the use of the CFG approach results in an increase in the amount of computation required. It appears, therefore, that there is no advantage gained, in terms of recognition performance and computational requirement, from the use of CFG approach over that of the FSG in the recognition of isolated words. The use of formal grammar approach over the direct storage of strings in isolated-word application makes possible the *generalisation' of strings in the training set. This can reduce the number of strings required by the learning process. Another advantage of the linguistic method is the reduction in the amount of computation in the FSG approach which is a result of the merging between similar segments of various strings during the training process.