Optimal filtering of noisy cepstral coefficients for robust ASR

In this work we investigate the use of a technique for optimal, in the mean-square-error sense, filtering of noisy cepstral coefficients for use with robust ASR. The filtering is done in the log-spectral domain using a clean speech model and a noise model whose parameters are estimated by a nonapproximative maximum likelihood formulation. As the assumption of additive noise in the time and spectral domains leads to a highly nonlinear mixing function in the log-spectral domain, we have to resort to numerical integration routines to perform the estimation and filtering. To make sure that the numerical integrals are robust and accurate we develop closed form solutions that cover critical parts of the integration domain.