A Research Of Voice And Complicated Background Noise Based On CNMF

Signal Source Separation(SSS), can be used as a key technology in many fields, such as bio-medical signal processing, image processing, speech processing and so on. In the area of speech processing, the mixing of target speech and background noise not only descends speech perception by human, but also dramatically degrade the performance of many speech processing technologies, for example the automatic speech recognition by machine.Therefore, speech separation, i.e., separating the target speech signals from the mixed signals is an important issue for speech processing.The fundamental knowledge of speech seperation skills and NMF are introduced.The existing NMF algorithms are divided into four categories in this thesis:(1) The tranditioanl NMF which only considers the nonnegativity constraint.(2) The NMF with some additional constraints as regularization.(3) The NMF with modified standard factorization formulation.(4) Some special NMF. And some brief introductions are given on basis of these respective methods in the end.This thesis deals with the separation of speech from the complicated background noise that appears in our daily lives. Convolutive non-negative matrix factorization(CNMF), which is improved version of the NMF, has recently been successfully applied in speech separation.In this thesis, a vector similarity measure is proposed to measure the similarity between the basic vectors from the CNMF output and the pre-trained clean speech basic matrix. By doing so, a new speech basic matrix containing speech-like vectors(based on the similarity measure)is rebuilt, with which the denoised speech signal is then reconstructed.The CHi ME database, which has been widely adopted for evaluating the speech separation algorithms, are adopted to evaluate the performance of the proposed algorithm, in comparison with the traditional NMF algorithms and the basic CNMF. Experimental results show that the proposed algorithm outperforms the NMF and CNMF.