A Research Of Speech Segmentation Method Based On Wearable Social Perceptual System

With the advancements of technology and the improvements of people's living,healthy issues has become the main problem in the current society.Usually,researchers can evaluate the physical and mental health objectively through social perceptual features.Speech signal processing is an extremely important research direction in this domain.This method can comprehensively assess physical and psychological health in social creatures.Thus,high-effective speech segmentation algorithm can be benefit for social speech perceptual features extracting.As boundedness of speech segmentation accuracy in HMM(Hidden Markov Model)method,which is one of traditional unsupervised speech segmentation algorithm,this thesis will present a novel HMM method mixing together with KL-Divergence(kullback-leibler divergence)and this new method can improve the accuracy in speech segmentation.But from the results of HMM-KLD method,there exists a problem of optimum decision strategy in segmentation algorithm.For this phenomenon,this thesis will propose the automated identification method of speech segmentation based on sparseness correlation feature and it can solve this difficulties effectively.This paper will be divided six parts and the specific content is as follows.(1)Based on research backgrounds about wearable social perceptual system and speech segmentation system,the thesis presents that high-effectively speech segmentation method can help social perceptual system analyze the speech features and be benefit for some research about the relationship between mental states and speech features in practical application.(2)Based on the traditional HMM unsupervised speech segmentation method,this thesis will evaluate the accuracy of speech segmentation.In this process,it will be devided three steps: step 1 is a process of denoising based on entropy and short-time correlation feature;step 2 is a process of removing the non-speech based on short-time energy;step 3 is the final process of segmenting the wearer's speech signal.(3)As the limitation of the HMM speech segmentation algorithm,this thesis will propose a novel HMM method mixing together with KL-Divergence and this method can improve the accuracy of speech segmentation.It also verifies the improvement effect of the new algorithm in the collected speech signal.(4)As the problem of the optimal decision in the HMM-KLD method,this thesis will propose a novel automatic speech segmentation method based on sparseness correlation feature.This method is a kind of automatic optimum decision strategy in speech segmentation algorithm and it can improve the accuracy of the final results in speech segmentation.(5)Based on basic speech features and rhythm speech features,this thesis will study the relationship between the speech features and the degree of intimacy in the conversation of different people,the social features in aged group.(6)In this part,it will summarize the superiority-inferiority in speech segmentation method and provide some suggestions.On the other hand,it also forecast the future development of speech segmentation method in the application of wearable social perceptual system.