Suppression of noise in speech using PARCOR analysis

Abstract

In this paper, a new noise suppression algorithm is presented for reducing any kind of background noise in speech. This noise suppressor uses three kinds of filters. The first filter, which is the inverse lattice filter implemented by PARCOR analysis of noise part, makes any colored background noises into white noises. The second filter, which is the white noise suppressor using spectral subtraction, reduces the white noise components in speech. The third filter, which is the noninverse type of the first one, restores the speech distorted by the first filter. The algorithm which detects nonspeech periods is combined with this noise suppressor, and the characteristics of these filters adaptively follow the nonstationary noise environment. The traffic noise and jet engine noise, etc., were used to estimate this system. As a result of the tests, we could improve the speech to noise ratio more than 20 dB. And at the same time, we could improve the articulation of the speech.