Double Gaussian based feature normalization for robust speech recognition

In this paper, a new feature normalization approach, based on the cumulative density function (CDF) matching principle, is proposed. Since speech features in noisy environments usually follow bimodal distributions, we fully utilize this characteristic by representing the CDF of the features with a double Gaussian model. A feature normalization process is performed according to the estimated CDF. The experimental results on the Aurora2 database show that the performance of our method is much better than that of the conventional mean and variance normalization (MVN) method, and comparable to that of the method combining spectral subtraction and histogram equalization (HE). Moreover, further improvement has been gained by combining our method with a simple temporal feature smoothing process. This result suggests that our new method has the potential to be integrated with other techniques to provide even better performance.