4.7.1 Word Graph Rescoring: Overview
![]() The word-graph incorporates a large number of hypotheses and defines which words can follow which words. It then can be used as the grammar constraints in a second pass decoding, which is typically referred to as the word-graph rescoring. Because the search space is strictly constrained at the word level by the word-graph, word-graph rescoring is much more efficient than the N-gram decoding. Although the process of word-graph generation is time-consuming, once a word-graph is generated, it can be rescored using more complex acoustic models and language models. This will give an overall better performance in terms of both speed and accuracy. For both one-pass N-gram decoding and word-graph generation, memory issue is a big problem towards efficient decoding. |
![]() ![]() ![]() ![]() ![]() |