Stochastic Segment Model Decoding Algorithm Based on Neighboring Segments and its Application in LVCSR

In the large vocabulary continuous speech recognition system based on stochastic segment model (SSM), the multistage decoding and pruning algorithm could decrease decoding time obviously. Generally, we only decode and prune for one segment each time. In this paper, a decoding algorithm based on neighboring segments is proposed. This algorithm decodes for multi-segments at the same time, so that the threshold of every segment could be highly shared by all the segments in each stage. That means more useless computation would be avoided, and the decoding would become faster. When using this algorithm in LVCSR system, we saved the decoding time of 50% approximately without accuracy loss.