摘要：De-noising of speech signal polluted by background noise, which is of great practical significance for effective transmission and accurate recognition of sound, is important. Local characteristic-scale decomposition is introduced, and it can divide the speech into low- and high-frequency parts without a loss in useful speech. The algorithm accelerates the speed of convergence, as well as pretreatment, and has less illusive component. In addition, two different criteria are proposed to select the reasonable noise reduction order based on the difference spectra of singular values. The proposed approach has improved the problem of noise reduction order selected by experience in singular value decomposition and enhanced de-noising effects. Simulation experiments verify the validity of the algorithm from subjective and objective evaluations.
关键词：Local Characteristic-Scale Decomposition;Singular Value Decomposition;Difference Spectrum;Speech Enhancement
