Speech Emotion Recognition Research Based On Noise Robustness

With the fast development of information technology and increasing requirement for human-computer interaction(HCI)technology,the demand for the emotional intelligence of machines is increasingly strong.As a necessary basis and precondition of the emotional intelligence of machines,automatic emotion recognition has theoretical significance and wide application prospect.As an efficient way of HCI,human speech carries rich information related to emotional state of the speaker,and the multidisciplinary issue of speech emotion recognition is attracting increasing research interest from various fields.However,there always exists a variety of environmental noises in the actual application of speech emotion recognition technology.For the task of speech emotion recognition,extracting and selecting noise robustness speech emotion features that helps to efficiently characterize different emotions,and constructing speech emotion recognition classifier with noise robustness are the main points of this thesis.In this thesis,this paper introduced the research backgrounds and importance of speech emotion recognition under the noise environment firstly,together with a review of related works.Key issues in speech emotion recognition research under the noise environment are discussed.Based on the construction of the optimal wavelet packet basis,the method of combination of short time frame and long time frame analysis and the sub-band spectral centroids of wavelet packet cepstrum coefficients,which is robust to the additive noise.This paper proposed a kind of spectral centroids weighted noise robustness wavelet packet cepstral coefficient for the speech emotion recognition task under the noise environment.At the same time,based on the speech segment trajectory model,this paper proposed the quantization criterion function for calculating the emotional information of the speech features to select the features from the high dimension spectral centroids weighted noise robustness wavelet packet cepstral coefficient features.Aiming at the environment noise,which is prevalent in test samples in speech emotion recognition,this paper proposed a noise robustness support vector machine classifier by weighting the relaxation variables in the equivalent optimization problem of support vector machine classifier.Experimental results at various levels of signal-to-noise ratio(SNR)show that the improved feature has better performance than.Finally,the feature fusion algorithm based on DBN,which combines the conventional acoustic features and spectral centroids weighted noise robustness wavelet packet cepstral coefficient,is proposed.With importance weighted support vector machine adopted as speech emotion classifier,effectiveness of the proposed feature fusion method in speech emotion recognition under the noise environment is verified by a series of comparative experiments.