Ensemble And Re-Ranking Based On Language Models To Improve ASR

We propose a strategy to improve speech recognition by selecting appropriate words to form new sentences using ensemble learning. Use traditional speech recognition methods first, and then rescore using different neural language models. Second, the decoding results of five different rescoring models were selected to select words. The choice of words is based on the importance of each word and whether the position of the word is correct. In the selection of word importance, the judgment method is majority weight and cumulative weight, and shift alignment and longest common subsequence alignment are used to determine word positions. And then the selected word representatives are reorganized to create new sentences. We compare the results of sentence ensemble and rescoring. As shown in the Aishell-1 test data, an error reduction rate of 7.30% can be achieved, verifying the effectiveness of the proposed method.