Topic Classification Of Speech Documents Based On Confusion Network

With the development of network and computer technology, the speech information dramatically increase in the amount of data, according to manage and make use of the speech information, spoken document classification is paid more attention in information retrieval, information filtering and information management applications area.A new method of topic classification based on confusion network was proposed in this paper. It can improve the ability of error correction by extracting the word information in confusion network. After post-processing of confusion network, the word information was combined with latent semantic analysis technology for improving the classification performance of the system. The method of weight calculation was improved. Different from traditional classification system, it combined word posterior probability to the classification when calculating TF-IDF weight.Topic classification system based on confusion network consists of three subsystems: speech recognition system, pos-processing system and classification system. In speech recognition system, the speech signal was processed by HMM model and HTK tools. The speech signal changed into three kinds of results including One-best, N-best and Lattice. After comparing the advantages and disadvantages of these three results. Lattice was chosen by the reason that it has a large number of candidates. In post-processing system, confusion network is generated by the clustering algorithm based on Lattice. The word information which can represent the speech document was extracted from confusion network. In classification system, confusion network was combined with latent semantic analysis technology. Compared with traditional latent semantic analysis, the method of weight calculation was improved in classification system based on confusion network. Then the dimension of vector space model was reduced by SVD and NMF. The semantic space after dimension reduction was used for training classification model.At the end of the paper, the results of all spoken documents including 6 categories of 8703 documents were listed. After analysising the results and comparing the results with One-best and N-best, the conclusion can be that topic classification based on confusion network can classify spoken documents accurately, and the performance was better than the results of One-best and N-best.