Full-covariance model compensation for noise-robust speech recognition