Empirical Distributions of DFT-Domain Speech Coefficients Based on Estimated Speech Variances

We present a novel way to estimate the empirical distribution of clean speech spectral coefficients. Rather than computing the histogram of clean speech within a certain signal-to-noise ratio interval, we normalize the spectral coefficients on the square-root of the spectral variance estimated via recursive averaging, the decision-directed approach or temporal cepstrum smoothing. We show that estimated distributions depend significantly on the used spectral variance estimator. Further, if the speech spectral variance is estimated in noisy conditions, the resulting histograms exhibit heavier tails as compared to clean conditions. The cepstral variance estimation approach is shown to result in less heavy tails as compared to the decision-directed approach.