Distributed processing of finite state grammars for connected word speech recognition

Abstract

An algorithm is presented for performing real‐time connected word speech recognition under the constraint of a finite‐state grammar, and executed in a multiprocessor pattern recognition computer. Each processing element of the computer consists of a 40‐MIP one‐chip pattern recognition processor and its associated local memory. [S. Glinski et al., “The Graph Search Machine (GSM): A VLSI Architecture for Connected Speech Recognition and Other Applications,” Proc. IEEE 75 (9), 1172–1184 (1987)] Aspects of the algorithm discussed include: the choice of models for speech sound units, words, and grammars; the problems of partitioning the models across processing elements; and the problem of removing redundancy in the finite state grammar. Extensions of the algorithm to more powerful grammars is also discussed. The capacity of the recognizer (i.e., vocabulary size) as a function of grammatical complexity is presented. Results indicate that high‐accuracy medium vocabulary speaker‐trained connected word recognition is readily attainable.