In the paper, a large vocabulary realtime isolated word recognition system (DSPELL) is presented. Although the final goal of the project is to recognise words from a vocabulary whose size is of the order of 10–20 thousand, the present system is intended to perform real-time recognition on vocabulary subsets (cohorts) of up to 2 thousand words. The system is implemented on Elsag's multiprocessor EMMA† for a fast response. Basic features of the system are the use of subword units (diphones) for the acoustic measurements, and the derivation of synthetic symbolic word templates directly from the lexicon. The use of diphones makes the training session of DSPELL very convenient.