Research On Speech Automatic Retrieval Technology For Broadcast News

Speech retrieval means for the users’ queries, searching from the speechdatabase and rerurning the docunment which satisfy the query request. Astechnology advances and increasingly development of Internet, the multimediainformation which people have accessed has grown exponentially, in whichbroadcast news speech occupies a large proportion, and easy to get, hasimportant scientific value. Speech retrieval technology can help people quicklyfind what they need from the massive data, so the technology has tremendoussignificance and practical value.Speech retrieval technology is actually the effective integration of speechrecognition technology and information retrieval technology, however, combinedsimply with the two techniques will lead retrieval performance is heavilydependent on the accuracy of recognition system, so the researchers will focus onmulti-candidate formal recognition results. The common recognition form ofmultiple candidates contains N-best Lattice and confusion network, wherein theN-best not contain all candidate results, and Lattice structure containing moreredundant and not conducive to the establish the index, so the confusion networkis often to be used. In addition, on the choice of the recognition units, if weselect words as the basic unit, the OOV probem will not be avoided. So the studyof Chinese speech retrieval technique often use in lattice structure based onsyllable.In this thesis, the automatic speech recognition system use thecontext-dependent techniques to train acoustic model, analyzes the impact ofmodel parameter sharing strategy during training and the number of Gaussianmixture, and combine language model to establish a large vocabulary continuousspeech recognition system. Then use the vector space model on the One Bestrecognition results to complete the baseline speech retrieval system. Later, in theform of a grid containing multiple candidate results, generating confusionnetwork which has more compact structure and through improved vector spacemodel, build a speech retrieval system. In addition, in the selection ofrecognition units, if the word is selected, it can not solve the problem “Out OfVocabulary”, and the system’s recall rate is low, while on syllables unit, it willreduce the accuracy of retrieval system, so this thesis presents a multi-unitshybrid retrieval system based on the backward integration.Experimental results show that the use of context acoustic modelingtechniques can adapt to the changes of pronunciation, improve the accuracy and robustness of the recognition system. The improved vector space model canbetter reflect the feature vectors in the proportion of the document, it canimprove the retrieval precision of retrieval system; rather more the use ofmulti-units hybrid retrieval system can not only solve the problem of OOVfundamentally, but also makes the system accuracy and recall rat e reached a goodbalance.