A numerical approach for estimating optimal gain functions in single-channel DFT based speech enhancement

We treat the problem of finding minimum mean-square error (MMSE) spectral amplitude estimators for discrete Fourier transform (DFT) based single-channel noise suppression algorithms. Existing schemes derive gain functions analytically based on distributional assumptions with respect to the speech (and noise) DFT coefficients and on mathematically tractable distortion measures. In this paper we propose a methodology to estimate the MMSE gain functions directly from speech signals, without assuming that speech DFT coefficients follow a certain parametrized probability density function. Furthermore, the proposed scheme allows for estimation of MMSE gain functions for pdf/distortion measure combinations for which no analytical solutions are known. Simulation experiments where noisy speech is enhanced using the estimated gain functions show promising results. Specifically, the estimated gain functions perform better than standard schemes, as measured by a range of objective speech quality criteria.