Abstract. Large-vocabulary continuous-speech recognition (LVCR) speakerindependent systems which i n tegrate cross-word context dependent acoustic models and n-gram language models are di cult to parallelize because of their interwoven structure, large dynamic data structures, and complex object-oriented software design. This paper shows how retrospective decomposition can be achieved if a quantitative analysis is made of dynamic system behaviour. A design which accommodates unforeseen e ects and future modi cations is presented.