Improved Autocorrelation-based Noise robust speech recognition using kernel-based cross correlation and overestimation parameters

This paper proposes a new algorithm to consider cross correlation between noise and clean speech signal when autocorrelation-based features have been used for robust speech recognition. Also, an overestimation parameter has been inserted in clean speech autocorrelation estimation. We have also adopted the normalization of mean and variance of energy and cepstral parameters as an extra means of further improving the speech recognition rate. We recently proposed a new approach for Autocorrelation-based Noise Subtraction (ANS). This method did not consider any possible cross correlation between noise and the clean speech signal. In this paper we have tried to consider this term during the estimation of clean speech signal autocorrelation. Our results on the Aurora2 corpus have shown that the recognition rate, when the cross correlation term is considered, is improved. Furthermore, taking into account the overestimation parameter further improves the results.