Kalman filter using quantile based noise estimation for audio restoration

This paper addresses the issue of removal of broadband noise from audio recordings which are degraded by aging or limitations of the recording-reproduction mechanism. To achieve the noise reduction, the Kalman filter is applied to the digitized audio signals. Kalman filter is an optimum filter for minimizing the error variance between a measured signal and its estimation. This is a time domain method and is free from musical noise phenomena inherent in the popular spectral subtraction method which is a spectral domain method. Application of the Kalman filter requires the correct estimation of the measurement noise variance (background noise variance in audio). Normally noise variance is estimated from silent regions of the signal. In audio applications this would be the first one or two seconds of the recordings. But this method may not be practical in all cases and especially if the signal SNR is very low. Hence in this work, the method of quantile based noise variance estimation is employed for determining the background noise variance. Performances of the filter with the two methods (silent region noise estimation and quantile based noise estimation) for broadband noise reduction in degraded audio signals are contrasted.