Noise Reduction Based On Low-rank Matrix With Sparse Matrix Decomposition

Voice communication inevitably be affected by various environmental noise,seriously affecting the practical application of speech technology. Speechenhancement technology is an effective way to solve the noise pollution, but noisestrongly and noisy environments enhanced still not been effectively resolved becauseof the complexity of the real environment of voice. Robust principal componentanalysis (referred RPCA) proposed that many practical observations can be attributedto a low-rank component and a sparse component, we can recover the original datainformation from large noise and pollution data with a low-rank and sparse matrixdecomposition theory. In this paper, we can through low-rank component correspondsto background noise and sparse component corresponds to speech and Constraints onrelated parameters, present a new speech enhancement method based on low-rankand sparse matrix decomposition. In the proposed method, noisy signal is transformedinto time-frequency domain where background noise is assumed as low-rankcomponent and human speech is regarded as sparse component, decomposition of thematrix can be obtained clean speech spectrum. In this paper, we propose an effectiveoptimization algorithms to solve the separation problem with low-rank matrix andsparse matrix, making speech enhancement effect is more pronounced. Experimentalresults show that speech enhancement method can steadily obtain higher noisesuppression performance in noisy conditions, compared to many traditional methods.The speech enhancement method based on low-rank and sparse matrixdecomposition is a new speech enhancement method. Different from traditionalmethods in principle and working methods, the method is adaptable to all kinds ofnoise, it can directly estimate enhanced speech and do not need voice activitydetector for noise estimation, and easy to adjust parameters. This method is expectedto achieve a breakthrough speech enhancement technology. Research project willpromote the development of signal processing theory and application of speechtechnology, and have high significance and practical value.