A Single-channel Speech Enhancement Technology Based On Jointly Constrained Dictionary Learning

During the communication process,speech is inevitably affected by the surrounding environment in real life.The intelligibility and clarity of speech signals are seriously affected by background noise such as mechanical sounds,traffic whistle and others' voices in communication that reduces the quality of speech signals.Speech enhancement technology which is used to extract the speech signals from these "polluted" noisy signals as pure as possible.In the current traditional methods of digital signal processing,the speech enhancement algorithms are divided into single-channel speech enhancement method and microphone array speech enhancement method according to the number of microphones in the channel.The methods studied in the paprer are all single-channel speech enhancement algorithms.In recent years,speech enhancement algorithms based on sparse representation and dictionary learning have been widely researched and applied in the field of speech enhancement and speech separation.The characteristics of speech and noise are considered by the single-channel speech enhancement method based on joint dictionary learning,but there are limitations on the treatment of interference noise.To solve the problem of "cross-projection" in the thesis,the discriminative joint dictionary is constructed by the constraints of the new optimization function and a single-channel speech enhancement method based on the joint constrained double-layer dictionary is proposed.The main content and innovations include:(1)The background,significance and development status of single-channel speech enhancement technology are briefly summarized.Meanwhile,the basic principles of sparse representation,dictionary learning and the single-channel speech enhancement algorithm based on dictionary learning are elaborated in detail.When using a single-channel speech enhancement method based on joint dictionary learning,the characteristics of speech and noise were only considered in the method that it is easy to generate the phenomenon of "cross-projection" between the sub-dictionaries of the joint dictionary,which causes source confusion and affects enhancement effect of noisy speech.(2)To solve the "cross-projection" phenomenon of joint dictionary,a single-channel speech enhancement method with a new optimization function based on joint dictionary learning is briefly described.In this new optimization function,the sparse representation of noisy signal on the joint dictionary is constrained,the approximate error of the speech signal and noise signal projected on the corresponding sub-dictionary is controlled.The error of cross-projection is minimized and the appropriate factors are used to balance the weights of the constraints to train more distinguishable joint dictionary.When noisy speech is enhanced by the method of the discriminative joint dictionary learning,the speech signal can be more projected on the speech sub-dictionary in the joint dictionary without being affected by noise sub-dictionary.Thereby the quality and intelligibility of the enhanced speech could be higher,and the effect of speech enhancement is improved.(3)In order to further suppress the confusing part of the joint dictionary,a single-channel speech enhancement method based on joint constraint double-layer dictionary learning is proposed.In the training phase,the characteristic sub-dictionaries that describe the clean speech and noisy speech are firstly trained.Then,with the new optimization function of discriminative constraints and anti-substitution constraints,a double-layer joint dictionary is trained which the first layer expresses the separable components of the speech signal and noisy signal,and the second layer expresses the easily decomposed components of the speech signal and noisy signal.The constraint of the objective optimization function is used to reduce the occurrence of "cross-projection" phenomenon and the confusion of the signals in the joint dictionary.Therefore,the effect of speech enhancement is further improved.