Research On Speech Enhancement With Adaptive Dual Data Stream

Currently most of the speech enhancement methods are based on microphones,such as Wiener filtering,enhanced methods based on wavelet analysis,etc.However,microphone speech is susceptible to environmental noise,thus affecting the enhancement effect,the performance of these methods will decrease significantly in low SNR environment especially.In order to improve the above shortcomings,this paper proposes an adaptive dual data stream speech enhancement method.Firstly,the Gaussian Mixture Model(GMM)and the noisy speech GMM are used to construct the adaptive MMSE filter to make filtering and reduce the noise by using the convergence feature of the microphone voice and throat transmitter;Then,the voice of the throat transmitter is converted through the priori model to be merged with the filtered speech using adaptive weight,that solves the problem of poor filtering speech when the SNR is low.The main research work of the paper includes:1.Firstly,it summarizes the research background and practical significance of this paper,then introduces the development histories and research status of speech enhancement technology both at home and abroad.2.A brief introduction and analysis of several traditional microphone-based speech enhancement methods have given in this paper,such as Wiener filter,MMSE and wavelet threshold de-noising method.Then both the characteristics and the applications in speech enhancement technology of the non-air conduction speech have been introduced3.The paper proposes an adaptive dual data stream speech enhancement method and then implements the algorithm.First,we use the HTK toolkit to train the cepstrum model,including the dual-stream GMM of the clean speech and the noise GMM,and then calculates their transcoding GMM model and the double-data noisy speech GMM model to construct the Wiener filter with adaptive weighting processing to improve the accuracy of model recognition,to obtain better enhancements.Experiments show that the enhancement effect of this method is better than that of the traditional microphone speech enhancement method and the single data flow GMM enhancement method.4.Due to the performance of the adaptive dual data stream GMM enhancement method is stuffy in low SNR speech,therefore,this paper will use the prior model to converse throat microphone voice to repair its high frequency component,and then makes a adaptive weighted fusion with the filtering speech,which enhances the filtering speech again to improve the enhancement effect in low signal-to-noise ratio.The experimental results show that the fused speech has better intelligibility and naturalness,and the PESQ score is improved.