Speech Enhancement Using Gaussian Mixture Models, Explicit Bayesian Estimation and Wiener Filtering

Gaussian Mixture Models (GMMs) of power spectral densities of speech and noise are used with explicit Bayesian estimations in Wiener filtering of noisy speech. No assumption is made on the nature or stationarity of the noise. No voice activity detection (VAD) or any other means is employed to estimate the input SNR. The GMM mean vectors are used to form sets of over-determined system of equations whose solutions lead to the first estimates of speech and noise power spectra. The noise source is also identified and the input SNR estimated in this first step. These first estimates are then refined using approximate but explicit MMSE and MAP estimation formulations. The refined estimates are then used in a Wiener filter to reduce noise and enhance the noisy speech. The proposed schemes show good results. Nevertheless, it is shown that the MAP explicit solution, introduced here for the first time, reduces the computation time to less than one third with a slight higher improvement in SNR and PESQ score and also less distortion in comparison to the MMSE solution.