Research Of Retrieval System From Speech To Speech

With the rapid development of multimedia information, more and more speech data begin to appear in people's daily life, speech information retrieval technology came into being. In this type of speech information retrieval, queries can be entered via text or speech. In this paper, a system through the speech query to retrieve spoken document is studied, which is retrieval from speech to speech.The retrieval system studied in this paper can be divided into two parts, one is speech recognition system and the other is information retrieval system. For the speech recognition system, it is constructed by an open source tools called HTK and the Chinese syllable is the basic unit of this system in this paper. From the point of smoothing algorithm in language model, an improved Katz algorithm combined with the SGT and Katz smoothing algorithm is proposed to improve the recognition rate of the speech recognition system. For the information retrieval system, the most widely used retrieval technology called VSM is adopted, whose index is constructed by TF and IDF. And the average retrieval accuracy is compared when the speech recognition result is in the forms of One-best and Lattice. Also, the impact of acoustic score in the syllable Lattice on the accuracy of retrieval systems is studied.Experiments show that the correct rate of speech retrieval system depends largely on the accuracy of speech recognition system. Lattice-based speech retrieval system can reduce the impact of error rate of speech recognition system. Compared the situation of One-best, the average precision can be improved by about 5.54%in Lattice-based speech retrieval system.