A Nonnegative Low-rank And Sparse Matrix Decomposition Based Speech Enhancement Method

When The speech signal was polluted by a variety of background noise and even covered, the process that from the background noise efficient extracte pure speech signal as possible as, improve the quality of voice and suppress and reduce the noise interference technique is called speech enhancement techniques. Speech enhancement is mainly to suppress background noise and enhance quality and intelligibility of the voice polluted by noise. In many occasions, we all need speech enhancement, which is a very important in speech processing technology. In the past years there were a lot of the classic speech enhancement algorithms which have been proposed, such as spectral subtraction, subspace Algorithm, the method based on statistical model. The speech enhancement has been widely used. Thus that to seek an effective algorithm which can deal with the speech signal with noise is the study of the purity of the original speech signal has great significance. Due to the random disturbance, it is impossible that completely extracted pure speech signal. The ideal situation is that I wish that speech enhancement algorithm can not only improve the voice quality, but also can improve the intelligibility. But they often seldom meet together. So the main challenge of speech enhancement is to design an efficient algorithm, which can effectively restrain the noise on the premise of not obvious signal distortion.Under the environment of noise for speech signal enhancement is a more complex task, in this in this paper, a new speech enhancement method was put forward. This method is based on the principle of low-rank sparse decomposition of nonnegative constraints in strong noise environment to realize the speech enhancement. The method is derived from the recently proposed in the robustness of principal component analysis(RPCA) principle. The speech signal and noise signal are as to speech signal and noise signal amplitude spectrum from the nonnegative low-rank sparse decomposition and optimization algorithm. In this algorithm the low rank matrix is corresponding to the noise signal spectrogram, and sparse matrix is corresponding to speech signal spectra.The experimental results show that in strong noise environment this method is betterthan some traditional speech enhancement methods, with less residual noise and low voice distortion, etc.