Parallel processing is now an accepted method of providing the computational power and flexibility required for large vocabulary continuous speech recognition. The Centre for Speech Technology Research has developed a large vocabulary real time continuous speech recogniser which addresses the issues of the efficient parallelisation of the hidden Markov model (HMM) Viterbi speech recognition algorithm. This system, runs on a flexible distributed memory parallel processing platform capable of sustaining increased vocabularies and improved recognition algorithms. The purpose of this paper is to describe the efficient parallelisation of the recognition algorithm and the underlying architecture and associated developments.