Two methods for estimating noise amplitude spectral in non-stationary environments

摘要

Estimating the amplitude spectral of noise signal is a very important part in many noise reduction systems. The conventional voice activity detection (VAD)-based method updates the amplitude spectral estimate only in speech absence areas and fails to deal with non-stationary noise. To overcome this problem, this paper proposes two methods to estimate the noise amplitude spectral for non-stationary environments: One is an indirect method, which obtains the estimate of noise amplitude spectral using its relationship with noise power spectral, while the other is the minimum mean-square error (MMSE)-based estimator. The proposed estimators are based on that the speech and noise are both Gaussian distributed and can update the estimate of noise amplitude spectral during speech activity as well as absence periods. Objective evaluations using several measures show that the proposed two estimators for noise amplitude spectral performed significantly better than the VAD-based method in all the tested non-stationary noise conditions.