An Acceleration Method of Noise Estimation for the Abrupt Change of Noise Level

In order to improve the tracking ability of the traditional noise estimation method when noise level changes abruptly, an acceleration method for noise estimation is proposed based on the Minima Controlled Recursive Averaging (MCRA). First, the burst detection of power spectrum is performed. If a sudden change of power spectrum is detected, a hangover period with adaptive length is enabled. Then, in the hangover period, a Voice Activity Detection (VAD) method based on the logarithmic likelihood ratio, spectral entropy and average magnitude difference is employed to detect the presence of speech. Finally, the update decision for noise estimation is made with the assistance of the ratio between noise estimation and power spectrum minimum. The test under ITU-T G.160 shows that, if the noise level is stationary, the acceleration method has no effect on the speech enhancement performance; while the noise level changes abruptly, the convergence time is reduced obviously, and the musical noise in the convergence period is also removed effectively.