On the comparison of front-ends for robust speech recognition in car environments

In this paper we compare several front-ends for Automatic Speech Recognition systems operating under noise conditions. The analyzed front-ends are based on standard MFCC parameterizations and include methods to compensate the effect of the noise over the representation of the speech signal. Three different compensation methods are considered in this work: Cepstral Mean Normalization, Spectral Subtraction and a novel method that we propose, based on a Statistical Compensation of the noise in the logarithmic Filter Bank Output domain. The considered front-ends are evaluated with Automatic Speech Recognition experiments using speech acquired in car environments. Making use of the French VODIS speech database (recorded in several cars running in real traffic situations) we have carried out recognition experiments to compare the different front-ends. The results show that the front-end including the Statistical Compensation of the noise outperforms the other considered methods.