Suppression of clipping noise in observed speech based on spectral compensation with Gaussian mixture models and reference of clean speech

Abstract

In recent years, the development of communication system allows people to easily record and distribute their speech. However, in the speech recording, clipping noise degrades sound quality when the level of input signal is excessive for the maximum range of amplifier. In this case, it is necessary to suppress clipping noise in the observed speech for improving its sound quality. Although a linear prediction method has been conventionally proposed for suppressing clipping noise, it has a problem with degradation of the restoration performance by cumulating error when the speech includes a large amount of clipping noise. This paper describes a method for suppression of clipping noise in observed speech based on spectral compensation. In this method, the power spectral envelope of speech on each frame in the lower frequency band is noise suppressed to by using Gaussian Mixture Models (GMM), and the one in the higher frequency band is restored by referring to the clean speech. We carried out evaluation experiments with a speech quality, and confirmed the effectiveness of the proposed method toward the speech, which includes a large amount of clipping noise.