Intelligent Speech Dialogue System Based On RASA

Intelligent Voice Dialogue System(Voice Chat Robot)refers to the system that intelligently communicates with human beings in real time through voice.Intelligent voice dialogue system is the master of speech recognition,natural language understanding,dialogue management,natural language generation and speech synthesis technology.It is an important research field in the practical stage of speech recognition technology.Voice interactive intelligent dialog system can let users directly ask questions through natural language.The system can return accurate and accurate answers,and give users a more intelligent user experience.It is believed that in the near future,natural language will replace input devices and touch screen to become the most widely used.This paper builds the whole system based on open source framework Rasa,and designs and develops a stable system which combines algorithm and application for the purpose of product landing and online service.Firstly,aiming at the problem of poor robustness of intelligent voice dialogue system in noisy real environment,an improved convolutional denoising selfcoding neural network speech enhancement algorithm is proposed,and the algorithm module is used as the front-end processing module of the whole system to improve the accuracy of speech recognition.Secondly,aiming at the low accuracy of text categorization in conventional machine learning algorithms,an improved multi-window convolution neural network text categorization algorithm is proposed,which can improve the accuracy of text categorization by learning the characteristics of n-gram grammar relations in linguistics.Aiming at the low recognition rate of single named entity recognition model and the assumption that text intent and entity are independent of each other,a joint model of LSTM+CRF is proposed.A model can do both intent recognition and named entity extraction,which not only improves the efficiency,but also improves the accuracy.Aiming at the problem of weak ability to deal with complex scenes and poor user experience in question-and-answer dialogue system,a LSTM multi-round dialogue management model based on sequence prediction is proposed,which enables the system to deal with at least 10 rounds of dialogue.Finally,aiming at the specific application scenario of food and beverage information query,the system will connect with the Wechat public platform to achieve real landing.The final system can realize intention recognition,entity extraction and multi-round dialogue management.For the catering industry,we need to be able to answer questions accurately,but also to deal with the topic of chat,giving people a more real experience.