Consideration of correlation between noise and clean speech signals in autocorrelation-based robust speech recognition

This paper presents a new approach to consider the correlation of noise and clean speech signals, when an autocorrelation-based set of features are used. Autocorrelation-based features have recently been used in several cases for achieving robustness in automatic speech recognition (ASR) systems. Such methods usually consider the clean speech and noise to be uncorrelated. However, when some correlation between the clean speech and noise is possible, their assumptions might not be necessarily correct. We recently reported a robust set of features based on autocorrelation of the noisy signal, namely Autocorrelation-based Noise Subtraction (ANS). The correlation between the clean speech and noise was not considered in this approach either. In this paper, we try to consider this cross correlation term in the estimation of the clean speech signal autocorrelation. This new approach was tested on the Aurora 2 corpus and led to even better results in comparison to ANS.