Kansei Image Evaluation Of Timbre Of Conversational Agent Based On Machine Learning

In the era of intelligence,voice interaction becomes a new paradigm of interaction.Consumer expectations for voice products has elevated from basic functions for practical use to the perception level,and they look forward to evoking more emotional experiences.As the important information carrier of VUI interface,timbre design is valuable for emotional design and evaluation.Based on Kansei engineering,this paper applied machine learning to quantitatively study the perceptual emotion of the timbre of voice products.The main work content can be divided into four aspects:First,a semantic space for evaluating timbre was established.In the early stage,a large number of vocabularies used to describe timbre were collected through desktop research and interviews,and preliminary screening was performed by experts.To reduce the dimensionality of the semantic space,a card classification experiment followed by two phases of cluster analysis was carried out.Secondly,the study conducted the sensitivity measurement to obtain the final dimensions used to evaluate the emotion of the timbre.The voice samples and perceptual dimensions were composed and sent out to form a questionnaire.a series of analysis including item analysis,factor analysis,and hypothesis testing were conducted to questionnaire data to get the result.Third,a timbre perceptual image evaluation model was established using machine learning.The feature parameter MFCC extracted from the corpus data as inputs,the score obtained from the timbre sensitivity measurement as a label,the mapping relationship was constructed by a neural network model.Based on the results of the above experiments and analysis,the evaluation model can calculate the score of each sentiment word pair with speech input,and predict the emotional tendency and strength of the timbre to the listener.The main innovation of this project is to introduce the method of perceptual engineering to quantitatively study the emotional emotion of timbre,and use the objective and stable evaluation model obtained by machine learning.This model is helpful for evaluating the performance of the timbre of existing voice products at the perceptual level.Fourth,based on the model,the product of Voice Design—an voice assessment tool is designed.By investigating the market prospect of emotional design of voice products,the enterprise designers are defined as the main users and their user scenario analysis is carried out.Four functional blocks are including voice evaluation,voice recommendation,comparative analysis and personal center are combed out.The main interface prototype are accomplished.This product aims to help evaluating the perceptual performance of existing voice products,benefit enterprises effectively designing the audio image of voice products,and improving the users'emotional experience.