Compensation of Noise Effects for Robust Speech Recognition in Car Environments
Résumé
In this paper, we propose a novel method to compensate the effect of the noise for Automatic Speech Recognition in car environments. This method can be applied to recognizers using a standard MFCC front-end. We perform a channel-by-channel compensation of the noise effect in the filter-bank output domain. In a first stage, the parameters describing the noise are estimated and secondly, we estimate the expected value of the clean speech in a probabilistic framework. The compensated filter-bank outputs are then used to obtain a compensated version of the MFCC-based parameters representing the speech signal. Recognition experiments using the French VODIS database (recorded in several cars running in real traffic situations) have been carried out to test the proposed compensation method. The results show the capability of the proposed method for the compensation of the noise effect in car environments.