SURE-MSE speech enhancement for robust speech recognition

This paper presents a new approach to enhancing noisy (white Gaussian noise) speech signals for robust speech recognition. It is based on the minimization of an estimate of denoising MSE (known as SURE) and does not require any hypotheses on the original signal. The enhanced signal is obtained by thresholding coefficients in the DCT domain, with the parameters in the thresholding functions being specified through the minimization of the SURE. Thanks to a linear parametrization, this optimization is very cost-effective. This method also works well for non-white noise with a noise whitening processing before the optimization. We have performed automatic speech recognition tests on a subset of the AURORA 2 database, to compare our method with different denoising strategies. The results show that our method brings a substantial increase in recognition accuracy.