Recurrent Neural Network Language Model For Continuous Speech Recognition

Language model is an important part of the continuous speech recognition system to obtain the language knowledge. And it has a very important influence on the recognition performance. n-gram language model is the most widely used among the models. In recent years, the Deep Neural Network has been gradually developed, and brings a new round of breakthrough for the speech recognition. The neural network language model is an important technique, which has received many great achievements in these years. Among them, the recurrent neural network language model is the most representative one. The RNNLMs have superior performance in the existing language models, and have been widely used in speech recognition, machine translation, information retrieval and other natural language processing tasks. Therefore, the RNNLMs are paid a great attention. This thesis studies on the RNNLMs in speech recognition, including decoding algorithm, the dependency of long distance and language model adaptation. The main contents are described in details as follows:Aiming at the problem that the Lattice decoding algorithm is difficult to use the RNNLMs, we propose an N-best rescoring method in two-pass strategy to use RNNLMs. We use a large 4-gram and RNNLMs to re-rank the n-best list and then output the new best hypothesis. For better performance, we introduce the cache RNNLMs to improve the decode accuracy on test data, which can get a more accurate language model score during the recognition process. The experimental results show that the proposed method can improve the performance of recognition system on test set.We present an improved recurrent neural network language model based contextual word vectors. The original RNNLMs are still lack of long dependence due to the vanishing gradient problem. This method can reinforce the ability of learning long-distance information by adding contextual word vectors into the model with feature layer. The vectors are obtained from CBOW and Skip-gram model. The experimental results prove that the proposed method can improve the performance of RNNLMs, and achieve obvious improvements in the word-error-rate.Finally, the performance of language model is highly depend on the consistency of training and test corpus, we present an empirical research on model adaptation and propose a new language model adaptation framework based RNNLMs. We design a three-step adaptation approach which training the general background RNNLMs on the adaptation data in domain to turning model parameters. And we obtain the topic features of the in domain data from LDA model and add them to the model training, enhance the model ability on self-adaptation in different corpus. The experimental results show that our adaptation method has better ability in different corpus and improve the recognition results.