Propagation of Statistical Information Through Non-Linear Feature Extractions for Robust Speech Recognition

摘要

Automatic speech recognition systems often rely on statistical noise suppression methods to increase their recognition performance in non-stationary noisy environments. However, even with a good approximation of the noise power spectrum, the estimated clean signal contains residual noise along with artifacts introduced by speech estimation inaccuracies. In this paper, we show that this can be compensated by propagating a measure of the uncertainty of estimation through the feature extraction process and combining it with missing feature techniques directly in the feature domain.