Noise Reduction in the Time Domain Using ARMA Filtering

In this paper, noise reduction in the time domain is discussed using autoregressive moving average (ARMA) filtering on a sample-by-sample basis. A major motivation for this approach is to avoid artifacts such as musical tones which often appear in conventional block processing schemes. The coefficients calculation is decoupled from the filtering process itself and can thus be carried out in the time- or frequency domain. A specific example for an ARMA filter structure for noise reduction is a Wiener envelope filter (WEF), which is derived by autoregressive (AR) modeling of noisy and clean speech using linear prediction (LP). The estimation of the LP coefficients of clean speech can be based on short-term (ST) power spectral density (PSD) processing. The ARMA filter coefficients can be additionally modified to suit the frequency resolution of the human auditory system for perceptional improvement.