Toward Real-Time Continuous Speech Recognition

Abstract

Abstract : This final report describes research on real-time speech recognition. The authors developed, under other DARPA-funded contracts, a system for continuous speech recognition. BYBLOS, the BBN Continuous Speech Recognition System, consists of a general paradigm and several algorithms for high performance speech recognition. The goals of the real-time project have been: 1) Develop algorithms and hardware for near-real time speech recognition; and 2) Develop and assess techniques for using highly parallel machines for continuous speech recognition. The authors investigated the use of the Butterfly parallel processor for increasing the speed of speech recognition. They found that is was possible to achieve factors of 10 speedup with minimal effort. More effort was required to achieve high efficiency on larger machines with 64 or more processors. Implemented was a near-real-time demonstration of continuous speech recognition with a grammar.