Research On The Algorithms For A Priori SNR Estimation In Speech Enhancement Technology

Speech is the one of the most convenient ways for human–computer interaction or communication between individuals.However,speech signal is inevitably disturbed or damaged by external background noise when it is transferred and processed,which not only leads to worsening perception and understanding,but also greatly reduces the overall performance of the speech signal processing system.The purpose of speech enhancement is to suppress the background noise in order to enhance the robustness of speech processing system and improve the quality and intelligibility of noisy speech.In this paper,the research work mainly focus on the problem of a prior signal-to-noise ratio(SNR)estimation in speech enhancement technology.It is well known that the a prior SNR is one of the significant parameters in speech enhancement technology,the rationality and accuracy of the a prior SNR estimation will in a large extent affect the overall output performance of speech enhancement systems.On the basis of studying a series of a priori SNR estimation algorithms,the paper analyzed how the value of the smooth factor in decision directed(DD)approach effects the performance of the overall speech enhancement system.The paper proposed a new type of a priori SNR estimation algorithm by introducing the coupling factor combined two DD approach with different smooth factor and proved the advantages of the new algorithm by theory analysis and simulation experiment.The research and innovation work in this paper is mainly includes:Firstly,this paper introduced and analyzed the characteristics of the speech enhancement algorithms in transform domain as well as the influence of the a priori SNR on the overall output performance.And the advantages and the disadvantages of the existing various kinds of the a priori SNR estimation algorithms were also illustrated and analyzed.Secondly,the most commonly used decision directed algorithm for the a priori SNR estimation has been detailed analyzed and researched utilizing the log-likelihood ratio and amplitude distribution histogram,for studying the speech distortion and musical noise residue situation of speech enhancement system.The experimental result showed that the overall output performance of a speech enhancement algorithm is controlled by the value of the smoothing factor,and the optimal value of the smoothing factor can be determined by considering the musical noise residual amount and speech distortion degree.Thirdly,based on the research results of smoothing factor,a new approach were proposed by using a convex combination of two DD methods with different smoothing factors,for the purpose to obtain a modified a priori SNR estimate with small speech distortion and musical noise residual.The proportion of either DD method to the convex combination is balanced by a coupling factor,whose value is optimally computed based on the minimum mean square error operator(MMSE)criterion.Finally,the properties of our proposed methods were tested by simulation experiments and the result shows the superiority of our methods in different noise conditions and SNR levels.We also summarized the full paper and prospected the research direction in the future for the estimation problem of a priori SNR in the speech enhancement technology.