OpenMP를 이용한 대용량 연속어 음성인식기의 병렬 구현

Abstract

method is employed in the emission probability computation for good load balancing. In the Viterbi search, however, we partitioned the search network into independent subtrees in offline to reduce memory synchronization overhead. The test was conducted using Wall Street Journal 20k evaluation and test set. We achieved the speed-up of 330% by utilizing four threads. The recognizer runs about 4.8 times faster than the real-time requirement.