Noise Power Spectral Density Estimation on Highly Correlated Data

In this contribution the Minimum Statistics noise power spectral density estimator [1] is revised for the particular case of highly correlated data which is observed for example when framewise processing with considerable frame overlap is performed. For this special case the noise power estimator tends to underestimate the noise power. We identify the variance estimator in the Minimum Statistics approach of being the origin of the observed underestimation. The variance estimator controls the bias compensation which is necessary to infer the mean power from a minimum value. This estimator turns out to be biased when the data is correlated. We provide an expression that describes the bias and show that by exploiting this the noise power estimation can be improved.