Research On Statistical Language Model Of Large-Vocobulary Continuous Speech Recognition System

Speech recognition is an important topic in pattern recognition field, because its development will deeply influence the future of human-computer interface. Among speech recognition technologies, large vocabulary, speaker independent, continuous speech recognition system is the most difficult and challenging one. Although many large vocabulary speech recognition systems (LVSRS) have been built up in recent years, none of them is satisfying and good enough for real applications. So, further research in this area is necessary and of great importance.Language model is a mathematics model that describes inherent disciplines of natural language. Along with the development of Corpus Linguistics, corpus-based statistical language model turn out to be an attractive alternative to manually constructed linguistic grammar, and widely applied to many research areas of natural language processing.The aim of this thesis is to construct a word-based context Chinese language model. The construction principle, evaluation criterion and current main problems of language model are discussed. Firstly a large, comprehensively sorted Chinese Corpus is constructed. The context is preprocessed and detached by word. CMU_Cam_Toolkit is used to finish the training and evaluation of language model. And a lot of experiments have been done to improve the performance of language model, such as the selecting of corpus contexts, the algorithm of preprocessing, the vocabulary file, smoothing and the adjusting the parameter of the toolkit. And these improve the word correction rate of speech recognition system greatly. The grapheme-to-phoneme conversion is a very important module for language model. It's used to connect the language model and acoustic model of speech recognition system. This thesis proposes to sort the phoneme of polyphone characters to high-frequency and low-frequency phoneme, and this method can largely reduce the labor consuming and time consuming. This thesis also proposed a data structure that can largely reduce the searching time. This model can be used in whole-word...