In this paper we present a solution to the nonlinear spectral estimati
on problem for speech enhancement. We start from a rather simple stati
stical model (log-normal) for the short time spectral estimates of spe
ech and noise. By empirical data generation and curve fitting approach
es we are able to get explicit, though simple, expressions for the MMS
E estimator in function of input level and the model parameters for ea
ch frequency component. The great advantage of our approach is that it
has a sound theoretical foundation, is general by the choice of its p
arameters, and almost as simple to use as classical spectral subtracti
on. Moreover, using a neural network as function approximator, which i
s found to be the best for our curve fitting problem, other model base
d MMSE estimators can be readily implemented with the proposed approac
h.