Noise robust zerocrossing rate computation for audio signal classification

Abstract

Zero-crossing rate (ZCR) is one of the most important acoustic feature that has been widely used in voice activity detection, voiced/unvoiced speech classification, music/speech classification image processing, optics, biomedical engineering, radar and fluid mechanics. The conventional time-domain ZCR measurement is sensitive to nonstationary noise. In this paper, we present a noise robust zerocrossing rate computation method. The number of zerocrossings is computed in autocorrelation-domain rather than in time-domain. The accuracy of the proposed measurement method is evaluated using both sinusoidal and speech waveforms under different signal-to-noise ratios (SNRs). Experimental results show that the proposed ZCR measurement method achieves better accuracy than the conventional ZCR measurement methods in the literature.