Application Of Long Distance Language Model In Speech Recognition

Speech recognition includes acoustic model decoding and language model decoding.Language model is an important technique of speech recognition.In the traditional language model,the 2-gram model is used to obtain the score of the language model,and the semantic information is short.In this paper,we use the improved word activation force long distance language model to obtain the long distance of natural language and semantic information.Through experiments,we show that the long distance information in the speech recognition and decoding process is effective,but also to explore the important role of the long distance language model in speech recognition.In this paper,the main research work and innovations are described as follows:1.Research the interpolation smoothing of word activation force based long distance language model.The traditional model of word activation force model is improved,which makes it more consistent with the traditional n-gram language model in the probability form,and thus can be used to smooth interpolation with the traditional n-gram language model.The experimental results show that the improved word based language model not only preserves the relationship between the adjacent words,but also has a long range of global relations.2.Proposing the decoding algorithm based on long distance language modelThe long distance language model is applied to the decoding of speech recognition,and use the relation between the historical word and the current word to expand the traditional 2-gram language model.The combination method of long distance semantic information and adjacent information is discussed,and the influence factors of the distance and the posterior probability of the historical words are considered.The experiment has proved the important role of WAF.3.Research the lattice rescoring based on long distance language modelImprove the WAF into two parts of forward and backward,and were applied to the post-processing process of speech recognition.and use the long distance relation to expand the lattice information.The process of the forward and backward-forward and backward direction is realized respectively combining the characteristics of the word activation force and lattice itself.The experiment has showed the Word Activation Force can play an important role in lattice rescoring...