Research On Automatic Evaluation Of Voice Politeness In Service Industry

Politeness is a complex social and cultural phenomenon.In the service industry,the importance of politeness is self-evident.As a service provider,how to hold the right politeness in the right environment and how to grasp the different politeness degree,it is an important issue to be considered.The automatic evaluation technology can provide such a help,through the objective score of voice politeness,to assist the service provider to produce more appropriate polite voice.This paper is devoted to the research on the automatic assessment of speech politeness in Chinese service industry.(1)A service industry politeness speech corpus is set up.The corpus includes four levels: very-politeness,comparative-politeness,a-little-politeness and neutrality.Based on the two factors that affect the politeness level: the speaker's speech content and expression,and the pragmatic distance between the two sides of communication,50 service industry scenarios are designed for each politeness level.Through recording,segmenting and tagging,a polite speech corpus with 1566 sentences is finally established.(2)This paper explores the automatic evaluation method of speech politeness.Using the framework of emotion recognition,the framework of automatic evaluation of speech politeness is built through speech signal preprocessing,feature extraction,feature selection and model construction.(3)Using random forest algorithm to select 89 dimensional speech features extracted,two features that have the greatest impact on speech politeness are compared and analyzed: the F0 range of 20 th to 80 th percentile,and the 80 th percentile of F0.The top 10 features of contribution are divided into three groups: F0 features,voice features and intensity features,and an in-depth analysis is carried out on them.The results show that F0 is directly proportional to politeness degree,F0 range is directly proportional to politeness degree,and F0 slope is inversely proportional to politeness degree;In female,F0 is significantly higher than male,but the change range of F0 is smaller than male.The change of voice is in direct proportion to the politeness degree;It is always larger in female than in male,while the range of it is on the contrary.The decrease of loudness is inversely proportional to the politeness degree;It decreases in female,and is stable in male.(4)KNN,SVR and GBRT are used to evaluate speech politeness automatically.The results show that the performance of the model with feature selection is better than that without feature selection.Among KNN,SVR and GBRT,KNN gets the worst performance,SVR achieves the best results in the fitting degree,GBRT gets the best model accuracy and the most consistent correlation between the predicted value and the actual value.So GBRT model has the best comprehensive performance for the automatic evaluation of voice politeness in service industry.