Enhancement of speech in additive, locally stationary and colored noise, using linear prediction

Abstract

In this study, an algorithm is developed for enhancement of speech, degraded by additive, slightly time varying, colored, background noise. The procedure is based on modeling both the speech and noise using linear prediction. It is well-known that a clean speech can be successfully represented by an autoregressive (AR) model. If the background noise is white, then the AR signal and the additive white noise correspond to an autoregressive and moving average (ARMA) model where the moving average (MA) part represents the effect of the noise. By using this fact, it is suggested to model the speech and the additive noise as two linear predictive models. The additive colored noise is whitened by filtering the speech using the estimated noise model. Then, the filtered signal in additive white noise is modeled as an ARMA process where the MA part represents the effect of the additive white noise on speech. Finally, inverse filtering the signal in additive white noise by the MA part yields a speech signal with eliminated noise.