Topic indexing of spoken documents based on optimized N-best approach

Abstract

For topic indexing of spoken documents, the word error rate is hopefully decreased instead of the whole sentence error rate, so the center hypothesis among the N-best results is selected as the final output in speech recognition system. Then all spoken documents can be represented as vectors with high dimensions in vector space model, which can be combined with non-negative matrix factorization or singular value decomposition to map the vector space into semantic space. Experiment results show that optimized N-best approach is more suitable to the topic indexing system than one-best method. Combined with the non-negative matrix factorization, the correct topic indexing can achieve 98.1% in optimized N-best approach, which is 0.9% higher than the one-best approach under the same condition. When the semantic space is decreased to 10, there is about 11.1% difference between these two approaches. Furthermore, compared with singular value decomposition method, non-negative matrix factorization has the advantages of better performance, faster computation speed and less storage space.