A lexical-tree division-based approach to parallelizing a cross-word speech decoder for multi-core processors

We present a novel approach to parallelizing a lexical-tree based LVCSR decoding algorithm for multi-core desktop processors. The approach distributes the search space among the cores by dividing the lexical tree in a way that minimizes communication between cores. Synchronization and load balancing schemes for this approach are described. The parallel algorithm is benchmarked on a 5k-word Wall Street Journal task. The context-dependent triphone model baseline system achieves a WER of 8.4%. The algorithm is shown to achieve a speedup of 1.63 on an Intel Core 2 Duo processor, with an average CPU utilization of 86%. The results also show that increasing the width of pruning schemes improves the parallel speedup.