Spoken Dialogue System For Service Robots And Research About Language Model

With the development of speech recognition technology, it has been applied in various areas. For service robots, speech recognition technology is used mainly in spoken dialogue system. In this study, we designed a spoken dialogue system specific to "KeJia robot" and explored the implementation process of it. In addition, we also studied how to improve recurrent neural network language model using unsupervised word clustering.The research about spoken dialogue system for service robots mainly involves two parts:speech recognition and dialogue management. In terms of speech recognition, we first introduced the theory of speech recognition in details, and then the corpus we collected for the application of KeJia robot, we described the detail steps of training an acoustic model and did the testing and analysis on the corpus we collected, The experiments showed that context dependent triphone model has the highest word accuracy of98.39%and sentence accuracy of90.83%related to it. Devices on robot usually have a limited computing ability, so we took advantage of the states information that the robot provides to design the dynamic language model mechanism that can compress the search space when do speech recognition decoding. The experiments showed that the best performance of our dynamic language model based speech recognition module can get the highest sentence accuracy of87.95%, which is12.05%higher than the speech recognition module that don’t use dynamic language model. In terms of dialogue management, we used python to build the dialogue management framework based on hierarchical state machine. Then we introduced the Adding of Multi-model information and the Verification and validation mechanism. Finally we introduced the practical application of our dialogue management framework on the task of "cocktailparty" of "KeJia robot".In addition, we explored the application of word clustering on the recurrent neural network language model. Recurrent neural network language model has been proved to be one state of the art language model. Researchers found that adding part of speech tag information to the input layer of recurrent neural network language model can improve the model efficiently. However part of speech tag require hand-annotated data to train the tag model, which consumes a lot and the extra tagger also makes the model more complicated, we proposed adding the results of brown clustering, instead of part of speech tag information to the input layer of the recurrent network language model to solve the problem. In the Penn Treebank corpus, the relative improvement over the original recurrent neural network language model reaches8%-9%.