Abstrait

A noise reduction method with modified running spectrum filtering

Zhang Yuxin, Ding Yan

In this paper we proposed a modified running spectrum filtering(RSF) method for noise reduction. RSF is an efficient noise reduction method. Because conventional RSF used FIR filter with 240 orders for reducing noise twice, the calculation cost is very high. Thus, removing low-frequency components with a bandpass filter can reduce the additive noise in power spectra. On the other hand, the Cepstrum Mean subtraction Algorithm (CMS) method is used to reduce the remanent noise in whole frequency domain in logarithm spectra. CMS is simpler than RSF. The calculation cost is far lower than that of RSF. The experiment shows proposed method not only improves the recognition accuracy of automatic speech recognition, but also reduces the calculation cost of conventional RSF