Speech Enhancement Based On Sparse Representation And Joint Dictionary Learning

Speech is an important medium for the communication of information and language.However,the speech signals are always contaminated and degraded by various types of noises in the real life.The degraded speech not only has a serious impact on the quality of voice communication,but also result in the auditory disgust and fatigue.Therefore,it is very important to study effective speech enhancement technology.The goal of speech enhancement is to improve the speech quality and intelligiblility by suppressing and eliminating the interferences and noises in the degraded speech.The traditional speech enhancement algorithm can achieve better performance in the stationary noise environment,and the restraining ability of conventional speech denoising methods is limited when in the non-stationary noise and speech-like noise environment.Therefore,based on the sparse representation and dictionary learning theory,this paper proposed an improved discriminative dictionary learning method to improve speech enhancement performance.In this paper,we introduce a speech enhancement based on joint dictionary learning method firstly.This method is mainly divided into two steps: dictionary learning part and speech enhancement part.In the dictionary part,we train a composite dictionary consisting of the concatenation of the speech dictionary and the noise dictionary.In the speech enhancement part,we use LARC algorithm with a suitably chosen residual coherence threshold to realize the separation of the speech and the noise.In the simulation experiment,our proposed method are compared with the conventional methods to verify their effectiveness.The joint dictionary learning method does not consider the distinction between speech samples and noise samples.Dictionaries are trained separately in the dictionary learning step,so it is difficult to avoid the existence of relevant atoms in the dictionaries,which will result in source confusion in the sparse reconstruction step.Therefore,this paper proposed an improved discriminative dictionary learning method,based on the joint dictionary learning.The principle of the discriminative joint dictionary method is to make different source signals be represented sparsely by their corresponding dictionaries and suppress the sparse representation in the non-corresponding dictionaries.It is especially difficult to overcome the non-corresponding dictionaries interference.In this paper,a valid correlation constraint condition is set up to reduce the correlation between the two dictionaries.The simulation results show that the proposed method is better than the traditional joint dictionary learning method in terms of quality measures of speech,and the proposed method is more suitable for denoising non-stationary noise.