Bias compensation methods for minimum statistics noise power spectral density estimation

The Minimum Statistics noise power spectral density (psd) estimation approach is based on tracking minima of a short term power spectral density (psd) estimate in frequency subbands. Since the short term minimum power is always smaller than (or in trivial cases equal to) the mean power, the minimum noise power estimate is a biased estimate of the mean power. For an accurate noise power estimate this bias must be compensated.In this paper we review bias compensation methods for moving average and first-order recursive smoothed psd estimates. While for some cases exact expressions for the bias are available, approximations are required in general. We present approximations which allow an efficient computation and compensation of the bias. We discuss factors that influence the bias and show that the method is to some extent robust to variations of the signal statistics. Besides different smoothing methods, we discuss the effect of overlapping spectral analysis windows and of signal correlation.