Improved feature enhancement using temporal filtering in speech recognition

The difference between training and testing environments is the major reason of performance degradation of speech recognition. In this paper, to further decrease the mismatch, we apply temporal filtering, Auto-Regression and Moving-Average (ARMA) filtering or RelAtive SpecTrAl (RASTA) filtering, as a post-processor for the log-Energy dynamic Range Normalization-Cepstral Mean and Variance Normalization (ERN-CMVN) based speech features, referred to as [EC]-ARMA and [EC]-RASTA. From experimental results conducted on Aurora 2.0 database, the integrated approaches with temporal filtering are shown the best performance among the several integrated approaches.