Robust Feature Extraction Using Autocorrelation Domain for Noisy Speech Recognition

Previous research has found autocorrelation domain as an appropriate domain for signal and noise separation. This paper discusses a simple and effective method for decreasing the effect of noise on the autocorrelation of the clean signal. This could later be used in extracting mel cepstral parameters for speech recognition. Two different methods are proposed to deal with the effect of error introduced by considering speech and noise completely uncorrelated. The basic approach deals with reducing the effect of noise via estimation and subtraction of its effect from the noisy speech signal autocorrelation. In order to improve this method, we consider inserting a speech/noise cross correlation term into the equations used for the estimation of clean speech autocorrelation, using an estimate of it, found through Kernel method. Alternatively, we used an estimate of the cross correlation term using an averaging approach. A further improvement was obtained through introduction of an overestimation parameter in the basic method. We tested our proposed methods on the Aurora 2 task. The Basic method has shown considerable improvement over the standard features and some other robust autocorrelation-based features. The proposed techniques have further increased the robustness of the basic autocorrelation-based method.