Noi se Reduction Based on Modified Spectral Subtraction Method
Ekaterina Verteletskaya, Boris Šimák · 2011
A bstract— In this paper, we propose a method for enhancing of speech corrupted by broadband noise. The method is based on the spectral subtraction technique. Performance of spectral subtraction, its limitations, artifacts introduced by it, and spectral subtraction modifications for eliminating these artifacts are discussed in the paper in details. To eliminate the musical noise, one of the artifacts introduced by conventional spectral subtraction, we propose to implement reduced varying scaling factor of spectral subtraction, with a following application of weighted function. Weighting function, used in the proposed algorithm, attenuates frequency spectrum components lying outside identified formants regions. The algorithm effects a substantial reduction of the musical noise without significantly distorting the speech. Listening tests were performed to determinate the subjective quality and intelligibility of speech enhanced by our method. Proposed noise reduction algorithm is compared to conventional spectral subtraction based on SNR improvement introduced by them. Spectrograms of speech, enhanced by the proposed algorithm and other modified spectral subtraction algorithms, which show the algorithms performance and degree of speech distortion, are also presented in the paper.