Research On Chinese Speech Retrieval Technology Based On Word Fragment And Lattice

Speech retrieval is one of new applications of speech recognition, and its aim is to search correlative information in amount of spoken data and return the results according to the query which is given by the users. It can realize fast context-dependent searching by primarily indexing speech resources.The target of this paper is to study the strategies to optimize the performance of Chinese speech retrieval system.First, the framework and theory of speech retrieval system are generally introduced.Secondly, automatic speech recognition (ASR) which is the front part of the system is optimized. A new unit"word fragment"of language model is proposed to take full advantage of the Chinese linguistic information between adjacent syllables, and an efficient algorithm for word fragment selection is studied. After word fragment selection of the training data labeled by syllable, it can be used to train syllable/word-fragment language model. And the word language model and syllable/word-fragment language model can be made combined as hybrid language model. The result of recognition system with syllable/word-fragment language model is better than the one using syllable language model as well as the advantage of hybrid language model to word language model.Thirdly, whole online part of the speech retrieval system is introduced, including generation of the indexes from lattices, searching of query, verification method according to the posterior probability and the processing of keyword overlapping. A method is proposed to change the lattices with multi-syllable units to ones which only have units instead of single syllable. With the lattices after processing the indexes which can resolve the problem of Out-of-Vocabulary (OOV) is gained. Another method is proposed to remove redundant nodes with a little cost of keyword detection, which can reduce the size of indexes significantly and the time of searching. The keyword detection performances of systems which use the indexes of different language models are compared.Finally, the future of speech retrieval is discussed in the conclusion.