RUB 

Noise Power Estimation

Most noise reduction algorithms which operate in the spectral domain require an accurate estimate of the noise power spectral density in order to compute the enhanced speech signal. Besides the traditional voice activity detection based methods for noise power estimation our Minimum Statistics approach has received significant attention.

The Minimum Statistics method is based on tracking minima of a short term power estimate of the noisy signal in frequency bands over time. The Minimum Statistics approach is capable of updating the noise psd estimate also during speech activity and leads to less clipping of the speech signal as there are no hard speech/no speech decisions. As long as the time window for the minimum search samples the noise floor speech cannot leak into the noise power estimate.

Our research interest is the development of improved noise estimation techniques which provide a good balance between a small estimation error and good tracking of non-stationary noise.

The Figures below depict the magnitude-squared Discrete Fourier coefficients, the smoothed coefficients and a noise power estimate for a noisy speech signal in a single frequency bin vs. time. The noise power estimate was obtained by tracking spectral minima over time. From the plot on the left hand side we find that the minima are biased with respect to the mean. For an accurate noise power estimate this bias must be compensated. The Figure on the right hand side shows the noise power estimate after bias compensation. Note that the noise power estimator is capable to track non-stationary noise also during speech activity.

signal frames