Research On Application Of Sparse Representation And Feature Dictionary In Speech Enhancement

As an important research area,speech signal processing is closely related to the high-speed development of the information age.Due to the interference of noise and reverberation,in many speech signal processing systems,the target signal often contains noisy speech signal.The polluted speech has great damage to the structure and intelligibility of the speech signal itself.Therefore,it is necessary to use some technology to reduce noise in the receiving end.Speech enhancement is one of the main technologies to solve the problem of noise pollution in real environment.When speech signal is polluted by noise,speech enhancement can effectively improve speech quality and intelligibility through some specific processing.The effective suppression of non-stationary noise is always the focus of researchers.The speech enhancement method based on dictionary learning and sparse representation provides a new way to remove non-stationary noise.Different from the existing dictionary learning speech enhancement algorithm,this paper selects the amplitude spectrum and power spectrum characteristics of speech signal and noise signal as the input of dictionary learning,and applies them to the system framework of speech enhancement algorithm.The amplitude spectrum and power spectrum features of speech and noise signals are used in the training dictionary,and the noisy speech is represented sparsely in the training dictionary.Finally,the reconstructed speech will be recovered by two weighted signals,one of which is obtained by amplitude spectrum and the other by power spectrum.Based on the difference and connection of the two signals,the following briefly summarizes the ideas of the paper.Firstly,in the existing speech enhancement algorithms based on sparse representation and dictionary learning,only one feature signal is used to recover the final enhanced speech.Based on sparse representation and feature dictionary learning,this paper studies the speech enhancement algorithm under single channel,and proposes a two-way feature speech enhancement algorithm.On the one hand,it increases the flexibility of the system framework,on the other hand,it fully exploits the relationship between the amplitude spectrum and power spectrum of the speech signal and the speech enhancement,and applies the dictionary of the amplitude spectrum and power spectrum to the speech enhancement comprehensively,which improves the problem of the different adaptability of the speech enhancement algorithm of the characteristic signal in different noise environments.Secondly,a noise power spectrum estimation method based on sparse representation and dictionary learning is proposed.The signal representation ability of sparse representation and dictionary learning methods mainly depends on the structural strength of the signal,and does not require the signal to have stationary constraints.It overcomes the defect that the traditional noise power spectrum estimation method can not update the noise power spectrum values of adjacent frames in real time,and improves the tracking estimation ability of noise estimation in non-stationary noise environment.Finally,for the important prior information of noise type,in the speech enhancement preprocessing stage,the noise classification module is added,and the noise type is determined by neural network algorithm.There are two kinds of noise types in and out of diversity,which ensure the accuracy of several common noise types.According to the type of noise obtained,the paper designs the adaptive weight of speech reconstruction in the speech reconstruction stage.The weight of two-way signal weighted recovery changes adaptively according to the change of noise type,so the noise adaptability of the algorithm is improved.