Connected speech recognition on a multiple processor pipeline

The authors report on efforts to put a connected speech recognition application on the MARS hardware accelerator. The beam-search algorithm chosen for this purpose was decomposed on a functional basis, not on a data-partitioning basis as in other attempts. This allows for easy synchronization and static load balancing. Further flexibility is provided by retaining the hierarchical structure of the speech graph. The implementation is being tested on a 1000-word vocabulary and preprocessed input data