Design And Implemention Of Personalized Voice Customer Service System

Intelligent voice customer service is the focus of machine learning and artificial intelligence algorithm research.For companies,it is necessary to save manpower and keep solving high-frequency repetitive problems.For individuals,it is also hoped that anthropomorphic intelligent customer service can quickly solve practical problems.However,today's customer service system still has problems such as insufficient semantic understanding,lack of multi-round dialogue ability,and insufficient anthropomorphic and personalized voice dialogue.This paper designs a personalized intelligent voice customer service system based on framework models such as Rasa,React,Tacotron2,MelGAN,etc.,which solves the key issues such as inaccurate responses in a single round of dialogue,loss of important information in multiple rounds of dialogue,and weak voice response capabillities.The basic service logic of multiple rounds of dialogue in the vertical field of meal,and the ability to use characteristic human voices to complete personalized voice services,this article mainly includes five major modules,namely the front-end interaction module,the back-end transfer request processing module,and the semantic understanding of Rasa NLU Module,dialogue management module and speech synthesis module1.The front-end interaction module uses the latest front-end functional component technology React Hooks in the React framework,which can achieve good results in coupling and scalability,and achieve smooth,concise and easy-to-use requirements in user presentation.Using the Ts programming language to standardize the format,the interactive interface of the intelligent customer service robot is realized,and the main functions such as text input,voice input and conversion text,text output display,voice output display,and tone output selection are completed.2.The back-end transfer request module is constructed using the lightweight Python asynchronous coroutine package Aiohttp,which is responsible for receiving the text input data of the front-end,and interacting the data information with the NLU part of Rasa and the dialogue management part to obtain the text reply output.Then interact with the speech synthesis module to convert the output response of the text into the output of the speech waveform,and finally return the information to the front-end module.The whole is used as the core transit module to connect to other modules,which is light enough and easy to expand in design.3.The Rasa NLU module is responsible for the entity extraction and intent recognition of a single sentence.It is a key module for semantic understanding.It uses the pipeline construction method to layer functions,including word segmentation,Chinese word vector conversion,entity extractor,and intent recognizer.And many other parts,using machine learning methods to train high-accuracy entity extractor model and intent recognition model,and with synonym extraction and regularization extraction to optimize entity extraction,using Spacy Chinese word vector model to optimize Chinese word segmentation and Chinese Vectorization.At the same time,the backtracking strategy before time is adopted to increase the amount of information in this round of dialogue,and the information from the previous rounds of dialogue has been included to greatly improve the accuracy of entity extraction and intent recognition.The data enhancement strategy is used to solve the problem of insufficient training corpus.The action selection strategy realizes the rollback of questions under low confidence and makes the question and answer more smooth.4.The dialogue management module is the core module to achieve multiple rounds of dialogue.This article builds an Action server for the logical process:ing of the dialogue,and closely cooperates with the Rasa NLU module to use the slot to manage the key information of the dialogue context,and use the knowledge graph to manage the basic entity information,And specifically optimized the extraction of Chinese numbers into Arabic numbers,and has a certain degree of derivation ability for unknown key information,and can generate anthropomorphic and accurate responses with the response template.5.The speech synthesis module includes speech synthesis model Tacotron2,vocoder model MelGAN,Green algorithm and vocoder model WaveFlow,combined with speech cloning technology to transform the speech synthesis model encoder part,solving a large number of single-person speech corpus collection Difficulty,based on the speech model of multi-person speech training,the speech model of special timbre can be obtained through migration training with less single-person speech data.In this paper,the vocoder model is trained to select the best solution through comparison and analysis.The final solution meets the requirements of good voice quality and low time-consuming generation at the same time.independently completed the design and implementation of all modules of the system.I used part of the open source data set on the data set.I collected Zhou Xingchi's voice data and self-recorded my own audio data as voice data.The order text dialogue data was also directly generated by my simulation..The system as a whole has reached the target requirements,and finally verified the core functions of the system and the real-time interaction and system stability through functional tests and performance tests.