Speech Recognition on a Multiple Processor Pipeline

We describe a large-vocabulary speaker-independent continuous speech recognition system being implemented on MARS, a multiprocessor developed at AT&T Bell Laboratories [1,2]. MARS consists of 15 processing elements (PEs) connected by a full crossbar, and fits on a single board which plugs into a Sun-3 workstation. Each PE is a custom microprogrammable VLSI chip which has a highly parallel architecture and specialized functional units. PEs communicate with each other by passing messages through the crossbar. The programming paradigm is that of control pipelining, in which the PEs are connected in a pipeline, each managing a single data structure.