This paper investigates the interface between the HMM pattern matcher and language models for speech recognition. A number of parameters are found to be important to the optimal performance of speech recognition systems. We consider the selection of match factors between these stages in the recognition process and their relationship with the width of the beam search. We develop improvements to a hybrid language model consisting a probabilistic context-free grammar and a bigram able to process both grammatical and non-grammatical speech input. Finally, consideration is given to the development of an algorithm for dynamic variation of the beam search pruning factor based on the ambiguity of the data input.