Research On Single-channel Blind Speech Separation Technology Based On Joint Dictionary

In many real-life scenarios,the speech signals we receive are often contaminated various sources of interference and environmental noise in the space.In severe cases,speech quality and intelligibility may be significantly reduced,resulting in poor human auditory perception and performance of automatic speech recognition.Therefore,in a variety of voice communication applications,such as hearing aids,speakerphones,and smart conference systems,it is desirable to utilize blind source separation techniques to efficiently recover source signals of interest from one or more observed signals.Among them,the single-channel blind speech separation problem with only one mixed signal and more than one source signal has always been an arduous task.The single-channel blind speech separation problem will be studied in the thesis.In recent years,after the theory of compressed sensing is proposed,sparse representation and dictionary learning theory have been widely used in the field of signal processing.Especially the single-channel blind source separation problem using dictionary learning has attracted more and more attention.The dictionary learning for single-channel blind speech separation usually uses the uniqueness of different source signals to achieve the separation of mixed speech,which has certain limitations.With the sparse representation and dictionary learning theory,the personality and commonality of source signals are considered.A single-channel speaker mixed speech separation algorithm based on an optimization function learning joint dictionary and a single-channel speech denoising algorithm based on the two-layer dictionary learning are proposed.Related experiments show that compared with the existing algorithms,the proposed algorithms works much better than the current algorithms.To sum up,the main research content of this paper includes:(1)The significance and development history of the blind source separation are introduced,and the recent development of the blind source separation technology is described.The concepts and the commonly used algorithms of sparse representation and dictionary learning are detailed.Then the framework and implementation process of solving the single-channel blind source separation problem using dictionary model are introduced.It is found that the separation of mixed speech can be effected by cross-projection only using the uniqueness of speech.(2)Because of the poor distinguishing ability of joint dictionary in single-channel speaker mixed speech separation using dictionary learning,the cross-projection engenders when mixed speech signal is represented on joint dictionary,which results in bad separation performance.In order to solve this problem,a new optimization function is proposed,and a speech separation algorithm based on this optimization function learning joint dictionary is put forward.In this algorithm,the identity sub-dictionary and the common sub-dictionary in the joint dictionary can be simultaneously trained.The identity sub-dictionary corresponding to the source signal represents its unique component,and the common sub-dictionary represents similar components of different source signals.Using the correlation between the sub-dictionaries,the optimization function ensures that the similar components of the source signal can be represented by the common sub-dictionary instead of the interfering sub-dictionary,which effectively suppresses the cross-projection.The experimental results verify that the proposed algorithm has better separation effect than the existing algorithms.(3)Aiming at the influence of confusing components in single-layer dictionary learning algorithm,a single-channel speech denoising algorithm based on two-layer dictionary learning is proposed.In the first level of dictionary learning,the existing discriminative single-layer dictionary learning method is used,but it is not required to completely suppress the interference between the speech signal and the interference signal in the layer learning.In the second level of dictionary learning,the confusion component generated in the previous layer of learning is suppressed by the constraint of the objective function.Through two-layer dictionary learning,the confusing components can be projected on the joint dictionary of the second layer,which reduces the interference between the speech signal and the interference signal in the single-layer learning dictionary,thereby improving the quality of the speech.The simulation results show that the proposed method has better noise reduction effect.