Estimators of noise. A. Smith.

We study estimators of the noise power spectral density. In comparable
settings the bias stays below 1% for all segment lengths we tried, and it
decreases as the number of averaged segments grows.
